Norm Autonomous Agents - CiteSeerX

Norm Autonomous Agents

Henricus J.E. Verhagen Department of Computer and Systems Sciences The Royal Institute of Technology and Stockholm University Sweden

May 2000

PhD Thesis ISBN 91-7265-673-7 DSV Report 00-004

c 2000 Henricus Verhagen

Printed by Akademitryck AB, Edsbruk, Sweden

Abstract Since the beginning of multiagent systems research, it has been argued that theories from the realm of the social sciences can be of help when building multiagent systems, and to some extent vice versa. This study sketches the concepts necessary for agents based on social theories. The concepts, viz. agent, autonomy and norms, used in these theories are discussed and defined. The social theories include theories on decision making, various rationality and action models, and the role and modeling of other agents. The sociological debate on the micro-macro problem is analyzed and translated to multiagent research and combined with philosophical theories on sociality. An agent typology is proposed and linked to the communication and modeling possibilities of the agents. Concluding the thesis is a selection of articles presented at workshops and conferences, focusing on the usefulness of norms for artificial adjustable autonomous agents, and two articles on simulation studies in order to develop and test organization theories.

Sammanfattning ¨ Anda sedan fleragentforskningens begynnelse har forskare argumenterat för att teorier fran ˙ de sociala vetenskaperna kan vara behjälpliga vid byggandet av fleragentsystem och, i viss man, ˙ ocksa˙ vice versa. Avhandlingen behandlar de begrepp som behövs för agenter konstruerade i enlighet med socialvetenskapliga teorier. Begreppen (mer precist agent, autonomi och norm) som används i dessa teorier beskrivs och definieras. Teorierna omfattar teorier om beslutsfattande, olika rationalitets- och agerandemodeller, samt andra agenters roller och modellering av dem. Den sociologiska debatten om mikro/makro-problemet analyseras och görs användbar inom fleragentforskning, i kombination med mer filosofiska teorier om sociala fenomen. En agenttypologi föreslas, ˙ inbäddad i kommunikations- och modelleringsmöjligheter av agenterna. Till slut presenteras ett urval

artiklar som presenterats pa˙ olika workshops och konferenser, fokuserade pa˙ användbarheten av normer för artificiella agenter, samt tva˙ artiklar som behandlar simuleringsstudier, utförda för att utveckla och testa teorier om organisationer.

Foreword The articles presented in my licentiate thesis were: • “TASSCS: A Synthesis of Double-AISS and Plural-Soar” (coauthored with Michael Masuch) [122]. • “On Being Social: Degrees of Sociality and Models of Rationality in Relation to Multi-Agent Systems” (co-authored with Ruud Smit) [107]. • “Modeling Social Agents in a multiagent world” (co-authored with Ruud Smit) [123]. • “Multiagent Systems as Simulation Tools for Social Theory Testing” (co-authored with Ruud Smit) [124]. • “ACTS in Action: Sim-ACTS - a Simulation Model Based on the ACTS Theory” [120]. After successfully defending my licentiate thesis in June 1998, I have (co)authored the following articles: • “Social Intelligence as Norm Adaptation” (co-authored with Magnus Boman) [19]. • “Adjustable Autonomy, Norms and Pronouncers” (co-authored with Magnus Boman) [117]. • “Adjustable Autonomy, Delegation and Distribution of Decision Making” (co-authored with Johan Kummeneje) [119].

• “On the Learning of Norms” [116]. • “Norms Can Replace Plans” (co-authored with Magnus Boman) [118]. • “Norms for Artificial Agent Action in Markets” (co-authored with Magnus Boman, Lisa Brouwers, Karin Hansson, CarlGustaf Jansson and Johan Kummeneje) [14]. • “An Anticipatory Multi-Agent Architecture for Socially Rational Action” (co-authored with Magnus Boman, Paul Davidsson and Johan Kummeneje) [15] The introductionary part will define the main concepts and clarify the threads of research connecting the articles. A subset of the above mentioned articles is then included. Apart from my co-authors I want to thank the anonymous referees for their valuable advice. Magnus Boman was indispensable as a supervisor, although I look onto our relation more as one of cooperation and colleagues than one of supervision in the traditional meaning. Also thanks to my former colleague Jan van Dalen at the Rotterdam School of Management for his help on issues of statistics. Travel funds issued by the Erasmus Trustfund helped me during my time in Rotterdam, while my PhD studentship was paid for by the NUTEK project Agent-Based Systems: Methods and Algorithms, part of the PROMODIS program. Last, but not least, thanks to my family, friends and Anna for their support.

Contents I

Concepts and Theories

1

1 Introduction 2 Definitions 2.1 The Definition of an Agent . . . . . . . . 2.2 Definitions of Autonomy . . . . . . . . . 2.2.1 Types of Autonomy . . . . . . . . 2.2.2 Adjustable Autonomy . . . . . . 2.3 The Definition of Norms . . . . . . . . . 2.3.1 Norms in Social Theory . . . . . 2.3.2 Norms in Legal Theory . . . . . . 2.3.3 Multiagent Systems Research and 2.3.4 Working Definition of Norms . . . 2.3.5 Learning of Norms . . . . . . . .

3

. . . . . . . . . . . . . . . . . . . . . . . . . . . . Norms . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

5 6 7 9 10 13 14 16 16 18 18

3 Internals of an Agent 21 3.1 Models of Decision Making . . . . . . . . . . . . . . . . 21 3.2 Considering Other Agents . . . . . . . . . . . . . . . . 23 3.3 Modeling Other Agents . . . . . . . . . . . . . . . . . . 25 4 Sociological Theories 27 4.1 Habermas’ Theory of Communicative Action . . . . . . 27 4.2 The Micro-Macro Problem . . . . . . . . . . . . . . . . 29 5 Levels of Agenthood 33 5.1 Dennett’s Ladder of Personhood . . . . . . . . . . . . . 33

5.2 5.3 5.4 5.5

Carley and Newell’s Model Social Agent . Conte and Castelfranchi’s Agent Typology Werner’s Agent Typology . . . . . . . . . Levels of Sociality . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

34 38 39 40

6 Proposed Agent Typology

45

7 Multiagent Systems Research 7.1 Reactive Agents . . . . . . . 7.2 Intentional Agents . . . . . 7.3 Believable Agents . . . . . . 7.4 Artificial Social Agents . . .

51 51 52 52 54

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

II Norms, Adjustable Autonomy and Simulations of the Spreading of Norms 55 8 Norms Can Replace Plans 8.1 Introduction . . . . . . . . . . . . 8.2 Adjustable Autonomy . . . . . . 8.3 Norms . . . . . . . . . . . . . . . 8.4 Conclusions and Future Research

. . . .

9 Adjustable Autonomy, Norms and 9.1 Introduction . . . . . . . . . . . . 9.2 Simulation of Learning of Norms 9.3 Pronouncers . . . . . . . . . . . . 9.4 Conclusions . . . . . . . . . . . .

Pronouncers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

57 57 58 61 62

. . . .

63 63 64 66 69

10 Adjustable Autonomy, Delegation and Distribution of Decision Making 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Simulation of Norm-Spreading and Norm-Internalization 10.2.1 Description of Decision Making Model . . . . . 10.2.2 Simulation Setups . . . . . . . . . . . . . . . . .

71 71 72 73 73

10.2.3 Implementation of the Simulation Model 10.2.4 Simulation Results . . . . . . . . . . . . 10.2.5 Interpretation of the Preliminary Results 10.3 Related Work . . . . . . . . . . . . . . . . . . . 10.4 Discussion and Future Research . . . . . . . . . 10.4.1 Simulation of norm-spreading and norm-internalization . . . . . . . . . . . 10.4.2 RoboCup . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

74 74 75 77 78

. . . . 78 . . . . 78

11 On the Learning of Norms 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 11.2 About Norms . . . . . . . . . . . . . . . . . . . . . 11.3 Simulation of Spreading and Internalizing of Norms 11.3.1 Description of Decision Making Model . . . 11.3.2 Simulation Setups . . . . . . . . . . . . . . . 11.4 Implementation of the Simulation Model . . . . . . 11.5 Simulation Results . . . . . . . . . . . . . . . . . . 11.6 Related Research . . . . . . . . . . . . . . . . . . . 11.7 Discussion and Future Research . . . . . . . . . . .

III

. . . . .

. . . . . . . . .

. . . . . . . . .

79 79 82 82 83 83 84 84 86 87

Simulations of Organizational Problem Solving

89

12 TASCCS: A Synthesis of Double-AISS and PluralSoar 12.1 Double-AISS . . . . . . . . . . . . . . . . . . . . . . . . 12.1.1 The Model of an Actor in Double-AISS . . . . . 12.1.2 The Search Space of an Actor in Double-AISS . 12.1.3 The Behavior of Double-AISS . . . . . . . . . . 12.1.4 Critique on Double-AISS . . . . . . . . . . . . . 12.2 Plural-Soar . . . . . . . . . . . . . . . . . . . . . . . . 12.2.1 The Model of an Actor in Plural-Soar . . . . . 12.2.2 The Actor’s Search Space in Plural-Soar . . . . 12.2.3 The Behavior of Plural-Soar . . . . . . . . . . .

91 91 92 93 95 95 96 97 98 98

12.2.4 Critique of Plural-Soar . . . . . . . . . . . . 12.3 Merging of Double-AISS and Plural-Soar: TASSCS 12.3.1 The Model of an Actor in TASCSS . . . . . 12.3.2 The Search Space of an Actor in TASCCS . 12.3.3 The Behavior of TASCCS . . . . . . . . . . 12.3.4 Results of TASCSS . . . . . . . . . . . . . . 12.4 Conclusions . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

99 100 100 103 105 106 109

13 ACTS in action: Sim-ACTS - a simulation model based on ACTS theory 111 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 111 13.2 ACTS theory . . . . . . . . . . . . . . . . . . . . . . . 112 13.3 Soar . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 13.4 Sim-ACTS . . . . . . . . . . . . . . . . . . . . . . . . . 115 13.4.1 History of Sim-ACTS . . . . . . . . . . . . . . . 115 13.4.2 The Agents in Sim-ACTS . . . . . . . . . . . . 116 13.4.3 The Warehouse Task . . . . . . . . . . . . . . . 117 13.5 The Results of the Simulations . . . . . . . . . . . . . . 118 13.6 Interpretation and Comparison of the Simulation Results119 13.6.1 Intra Model Comparison . . . . . . . . . . . . . 119 13.6.2 Inter Model Comparison . . . . . . . . . . . . . 120 13.7 Conclusions and Discussion . . . . . . . . . . . . . . . 121

List of Figures 5.1 The agent categorization matrix . . . . . . . . . . . . . 35 6.1 The proposed agent typology . . . . . . . . . . . . . . 47 10.1 10.2 10.3 10.4

The The The The

norm-sharing factor with autonomy = .4. . . . . norm-sharing factor with leadership = 2. . . . . norm-internalization factor with autonomy = .4. norm-internalization factor with leadership = 2.

. . . .

75 75 76 76

11.1 The normspreading factor for all simulation setups . . . 85 11.2 The norminternalizing factor for all simulation setups . 85 12.1 12.2 12.3 12.4 12.5 12.6 12.7 12.8

Graphical representation of the warehouse task . The warehouse task and its subtasks . . . . . . The main cycle of TASSCS . . . . . . . . . . . . The find subtask . . . . . . . . . . . . . . . . . The get subtask . . . . . . . . . . . . . . . . . . The put subtask . . . . . . . . . . . . . . . . . . The move problem space . . . . . . . . . . . . . The evaluation of a message . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

99 102 103 104 104 105 106 107

List of Tables 12.1 Summary of the results of TASSCS . . . . . . . . . . . 108 12.2 Summary of the results of Plural-Soar . . . . . . . . . . 108 13.1 Results of the simulation of two independent agents using the Sim-ACTS model. . . . . . . . . . . . . . . 13.2 Results of the simulation of two independent agents using the TASSCS model. . . . . . . . . . . . . . . . 13.3 Results of the simulation with agent X being egoistic and able to command agent Y who is altruistic using the Sim-ACTS model. . . . . . . . . . . . . . . . . . 13.4 Results of the simulation with agent X being egoistic and able to command agent Y who is altruistic using the TASSCS model. . . . . . . . . . . . . . . . . . . .

. 118 . 118

. 119

. 119

Part I Concepts and Theories

Chapter 1 Introduction The central problems of sociology and multiagent systems are closely related, almost complementary. Sociology, originating in the early days of the industrial revolution and the breakdown of traditional social ties, tries to find explanations for the existence of social order among collections of more and more free yet interdependent individuals. Multiagent system designers face the problem of how to ensure efficiency at the level of the system whilst respecting individual autonomy. In this thesis, social theories are called upon to point towards possible solutions of this multiagent system research problem. Even though the reader may sometimes frown upon the use of theories based on human social systems (or the idea of artificial social systems for that matter), it is my firm belief (a belief also put forward in [20] from the early days of multiagent system research) that these theories can be put to use in multiagent systems, and that the use of multiagent systems may be of great benefit for the development of the social sciences. In the introductionary part, several social theories will be presented and compared with respect to their view upon the individuals contained in the social system and the properties ascribed to the social system itself. One possible solution to the problem of combining social level efficiency with autonomous agents is the use of central control (thus limiting the individual’s autonomy severely). In human social systems such as organizations this is realized via bureaucracy. In multiagent

4


systems it is the central coordinator that plays this role. This solution works only when the social system’s environment has a low rate of change (this includes the set of individuals included in the social system) since central control has as one of its main characteristics a low rate of adaptability (see [29] for a discussion on the impossibility of an optimal organizational structure). When flexibility is of essence, other solutions are called for. An intermediate solution is internalized control, e.g. the use of social laws [102]. The most flexible system involves a set of norms and learning at all levels, including the level of norms, based on reflecting upon the results of actions. Structural coordination as proposed in [90] is another example of an intermediate solution that is only suitable for closed systems (or at least systems in which the behavior of new members has to conform to preconceived rules). Open systems [56] (with respect to the composition of the social system and the environment in which it is to function) require the most flexible solution. During the lifetime of the system, norms evolve (or possible even emerge) to adapt to changes in the circumstances in the physical and social world. The research presented in this thesis focuses on two closely connected concepts for artificial agents, viz. autonomy and norms. Both concepts will be dealt with in the first part of this thesis. I will also provide a definition of agency, sketch different models of decision making, the micro-macro problem will be introduced, and I will introduce various models of levels of agenthood as well as my own typology of agents based on levels of autonomy. Furthermore, levels of sociality will be addressed, as well as the role and modeling of other agents. To conclude this first part, various visions on social agents will be compared and linked to the levels of agency as distinguished in my typology of agents.

Chapter 2 Definitions I will attempt to combine theories from the social sciences (sociology, psychology, legal theory, and social philosophy) with theories developed in multiagent systems and some of the theories behind multiagent systems (such as cognitive science, economics, decision theory). Some of the terms used are defined differently within these different fields of science. This calls for descriptions and comparisons of the most relevant definitions and clear statements as to which definition is used in this thesis. An example is the term “normative reasoning”. In decision theory for example, this means “reasoning using normative arguments (norms being prescriptive and norm compliance obligatory)”, whereas in sociology this term means “reasoning about norms”. The latter definition is the one used in this thesis. The differences between these two definitions are grounded upon the different definitions of norms to which I will return below. Another concept that needs a (short) definition is emergence. In this thesis emergence expresses arising properties at the macro level that are not reducible to the properties at the micro level of the individual agents.1 In the remainder of this chapter, the central concepts used in this thesis will be discussed, viz. agents, autonomy, and norms. 1

For a more elaborate discussion of emergence, see [63].

6

2.1


The Definition of an Agent

Before starting with multiagent models, or more precisely models of decision making by agents in a multiagent environment, it makes sense to look at models of single agents, although the term agent makes little sense in a single agent world. Such a model follows the solipsistic paradigm, i.e., in the deliberation process of an agent there will be no account of other agents simply because they do not exist. This is the line of work that cognitive science has focused on. To quote Werner [128] “cognitive science is fundamentally an asocial science.” One main problem with the use of the term agent is apparent by just looking at the dictionary definitions of the word “agent”. One meaning is “one that acts or exerts power”, whereas another is “something that produces or is capable of producing an effect” or even stronger “a means or instrument by which a guiding intelligence achieves a result”. The first definition underlines the pro-activeness, the other definitions express the limited autonomy of agents and the possibility of agents to change their environment. It is this duality that lies at the bottom of the confusion about the definitions of “agent”. As a consequence of this ambiguity, it is not surprising that the definition of an agent varies between different multiagent systems researchers (cf., [130], [55] for definitions and listings of definitions). In [129], the basic set of agent properties consists of autonomy (i.e., the agent has some kind of control over its internal state and actions), social ability (i.e., the ability to interact with other agents which may include humans), reactivity (i.e., the ability to act upon perceptions of the environment of the agent) and pro-activeness (i.e., the ability to display goal-directed behavior so that the agent’s behavior is not only ruled by environmental changes or social interaction). The ability of an agent to interact with its environment in pursuit of its own goals is in [109] labeled automaticity or control autonomy. In [55], an autonomous agent is defined as a system situated within and part of an environment that senses its environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in

Definitions

7

the future. This definition more or less implies that there also exist nonautonomous agents, but the authors fail to give a definition of these agents. Most definitions of agency used in multiagent systems research do share one main characteristic, namely that an agent is an entity that acts autonomously. This places the definition of agency in the midst of philosophical debate on autonomy as well as action, just like AI plunged into the philosophical debate on intelligence and consciousness. This definition also turns expressions like autonomous agents into tautologies (although according to [42] we might reserve that term for agents that are both situated, i.e., are part of a society, and embodied, i.e., have a physical body that can move in the real world, e.g., a mobile robot). The properties of agents can be used to characterize different types of agents as well as the agents’ and system’s possibilities and limitations [33], [123], [124]. The research goal can thus be used to decide on the type of agent and system to be implemented [31], [30].

2.2

Definitions of Autonomy

Both within and outside multiagent research there are several definitions of autonomy. I will take as a starting point the definitions of autonomy and autonomous as found in the Webster dictionary.2 autonomy: 1. the quality or state of being self-governing; especially: the right of self-government 2. self-directing freedom and especially moral independence 3. a self-governing state 2

The Webster dictionary can be found on-line at www.m-w.com/dictionary.htm.

8

Concepts and Theories autonomous: 1. of, relating to, or marked by autonomy 2. (a) having the right or power of self-government (b) undertaken or carried on without outside control (example: an autonomous school system) 3. (a) existing or capable of existing independently (example: an autonomous zooid) (b) responding, reacting, or developing independently of the whole (example: an autonomous growth) 4. controlled by the autonomic nervous system

The first definition of autonomous is more or less circular when combined with the definition of autonomy, as is definition 2a. The definition given in 2b is the degree to which decisions are directly determined by external stimuli (autonomy from stimuli in [35]), in other words the degree to which the behavior of an agent is directly controlled by another entity at runtime.3 Another interpretation of definition 2b is the independence of the human mind: cognitive autonomy, the degree to which one’s choices and actions based on them are governed by other agents (see also [35]). These definitions refer to autonomy at the level of the individual agent. At the level of a multiagent system, autonomy is the autonomy a society has to give itself laws (definition 2a). I will call this norm autonomy, i.e., the degree to which one’s choices (and the actions based on these choices) are governed by social influences. The autonomy to give oneself laws can also be applied at the level of the individual agent where it represents the learning capabilities of the agent. These learning capabilities may have the agent diverge from predefined decision making rules, thus changing its behavior. Autonomy of the agents 3

Note that the above cited definition of an agent by [130] uses both the term pro-activeness and the term autonomy to capture this definition of autonomous.

Definitions

9

forming a multiagent system with respect to its designer(s) is related to the self-organizing capabilities of a multiagent system. The more autonomous the agents are, the higher the multiagent systems selforganizing capabilities are, thereby making emergent behavior more feasible. Autonomy in this sense may be equaled to the control an agent has of different factors that play a role in its decision making process. I will return to this when I describe my typology of agents. Autonomy is relative to something external to the agent: an agent is not autonomous in an abstract sense but is autonomous on some level with respect to another entity, be it the environment, other agents, or even the developers of the agent. This point is also made in [35] where it is stated that an agent’s autonomy is necessarily limited, since it is situated. Unlimited autonomy renders an agent solipsistic since the agent does not allow for interaction with its environment. There are also various types and grades of autonomy. These will be discussed in the next section. 2.2.1

Types of Autonomy

During his invited talk at MAAMAW’99, Parunak [93] described two types of autonomy (in the sense that an agent decides by itself when it executes, in contrast to objects in OOP), viz. dynamic autonomy and deterministic autonomy. Dynamic autonomy is the ability to start execution, the pro-active side of agents which is determined by the agent’s internal structure. Deterministic autonomy is the ability to deny access to the agent’s resources, influencing the predictability of the agent. Strangely, Parunak states that this is determined by an external observer. This autonomy is then not autonomy of the agent but an indication that the external observer has an incomplete model of the agent’s internal model. The deterministic autonomy is however a useful concept if the control over access to its resource is managed by the agent.

10


Boden [11] (in the context of A-Life) defines the following three dimensions of autonomy as control over behavior: 1. the extent to which behavior is directly controlled by the environment 2. the extent to which the controlling mechanisms are selfgenerated (emerging) rather than externally imposed (deliberatively preconfigured). These include hierarchical levels of abstraction (groupings of individuals that can be considered as a unit themselves) 3. the extent to which inner directing mechanisms can be reflected upon and/or selectively modified Purely reactive agents are agents that have no inner mechanisms: external stimuli are directly coupled to behavioral output. Behavioral autonomy is complemented by operational autonomy or automaticity (being able to operate in an environment, sense that environment and be able to change it in accordance with one’s goals [109]). Operational autonomy is the engineering view on autonomy that we recognize from most work on adjustable autonomy. 2.2.2

Adjustable Autonomy

In [75], adjustable autonomy was first defined as integrated system management (in the space exploration domain) being “able to be overridden by the operators at any chosen level of shared or traded automatic control”. In [51], adjustable operations are defined more narrowly as “operations which are usually autonomous, but provide shared and/or traded control whenever the crew chooses. Crewcentered adjustable autonomy is defined as giving the crew user of the system control over and improved situational awareness insight into the current, past (insight only), and possible future operations of the system whenever the user chooses.” The general idea is that autonomous systems adapt the level of autonomy to the situation

Definitions

11

at hand. Individual agents participate in several groups and may revise their standing toward the group and its members during execution. This enhances their self-organizing capabilities and increases their flexibility and (re)action repertoire. In a more recent report [47], the level of autonomy is defined in this narrow manner, viz. as being dependent on: • the complexity of the commands • the number of autonomously controlled subsystems and couplings between these systems • the circumstances that will allow for override of manual control • the circumstances under which user information or control is requested (1999 IJCAI Workshop on Adjustable Autonomous Systems Call for Papers) • the duration of autonomous operation (restated in 1999 IJCAI Workshop on Adjustable Autonomous Systems Call for Papers as the resources consumed, including time) An example of the work on adjustable autonomy is [5]. The autonomy level is measured per goal and is related to the amount of agents that decide what goal to pursue. Is the agent itself the only agent involved in making the decision, its autonomy value for that goal is equal to one. If the agent has no say whatsoever in which goal is to be pursued, its autonomy value is zero. Consensus (unanimity) lies in the middle of this autonomy spectrum, here the autonomy level is the reciprocal of the number of agents. Since the autonomy level is dependent upon the amount of agents, it is hard to compare the value over time and between different multiagent systems. Several authors (cf., e.g., [39], [33], [128], and [44]) in multiagent systems research have developed more or less overlapping hierarchies of autonomy based on the types of processes for decision making that the agents use. These autonomy models are summarized and extended

12


in [124]. In short, decision making takes place at four separate yet connected levels (see also chapter 6 below): • the level of actions • the level of plans • the level of goals • the level of norms These two definitions of levels of autonomy are quite different. The basic difference lies in their view on autonomy: • levels of autonomy as abstraction levels, or the control the agent has over its behavior and decision making process • level of autonomy as level of independence of group (or human supervision) The work on adjustable autonomy is concerned with the second type of autonomy whereas multiagent systems research is primarily concerned with the first. Reconciling these two types of autonomy is a non-trivial problem. The complexity of the commands (from humans to agents) can be measured in terms of the transfer of control (or delegation) that the communication yields in terms of the levels of actions, plans, goals, and norms. The following example is taken from [47]. Consider the commands: “find evidence of stratification in a rock” and “go straight 10 meters”. The first command transfers a goal to the agent, with the agent assuming responsibility to judge what is to count as an adequate result, which object to examine, etc. The second command also transfers a goal but also what is to count as an adequate result. It can even be debated whether this second command is at the level of actions or not. Autonomy in the social space increases with the number of subsystems under agent control since the agent is less dependent on other agents to perform tasks. This is related to the action repertoire of the agent. Note also that

Definitions

13

this is connected to the second type of behavioral autonomy as defined by Boden and described in the previous subsection. The need for, and implementation of, adjustable autonomy for governing overriding through human commands is well documented in the agent literature. One domain in which a multiagent system with relatively autonomous agents must expect frequent human intervention is intelligent buildings [18]. Agent intervention in human activity is less studied but occurs e.g. in surveillance, alarm, and tutoring systems [95]. Decision making at the level of norms (or in more simple cases applying norms) is to decide on overriding manual control or for that matter any outside control [12]. The interaction between the agent and human supervisors (e.g., requests for information) is not modeled as a separate level of autonomy in the agent theory literature. Interaction consisting of information transfer from the human supervisor to the agent is conceptually equivalent to results of information gathering from the environment. Requests for transfer of control is equivalent to the transfer of control stipulated above. The introduction of time as a separate autonomy level (the last point in the list of autonomy levels in the adjustable autonomy literature) is an important addition, although it in fact is merely making explicit the role time plays in the multiagent systems view on autonomy. The levels of decision making are closely linked to time. Both the time-span of execution and the time-span used for self-reflection increase with the levels mentioned. Having control over decision making and behavior over increasing periods of time makes the agents less dependent and thus more autonomous.

2.3

The Definition of Norms

In this section I will first describe the various uses of the word “norm” in different scientific arenas before stating the view I use. After this I will take a closer look at the work on norms I have conducted for this thesis. But first let us once again take a look at the Webster online

14


dictionary: 1. an authoritative standard (model) 2. a principle of right action binding upon the members of a group and serving to guide, control, or regulate proper and acceptable behavior 3. average as: (a) a set standard of development or achievement usually derived from the average or median achievement of a large group (b) a pattern or trait taken to be typical in the behavior of a social group (c) a widespread practice, procedure, or custom (rule) These different views on norms are reflected in the different views on norms in different scientific disciplines. For my purposes I will describe the views on norms in legal theory, (social) psychology, (social) philosophy, sociology, and decision theory. Where appropriate I will also describe different views even within these disciplines. After the presentation of these different views, a working definition is presented and the learning of norms is addressed. 2.3.1

Norms in Social Theory

In the description of the normative action model, Habermas [65] identifies the use of norms in human action patterns as normatively regulated action. The central concept of complying with a norm means fulfilling a generalized expectation of behavior. The latter does not have the cognitive sense of expecting a predicted event, but the normative sense that members are entitled to expect a certain behavior. This normative model of

Definitions

15

action lies behind the role theory that is widespread in sociology ([65] p.85, original emphasis). This view is in agreement with Tuomela [113], who distinguishes two kinds of social norms (meaning community norms), viz. rules (rnorms) and proper social norms (s-norms). Rules are norms created by an authority structure and are always based on agreement-making. Proper social norms are based on mutual belief. Rules can be formal, in which case they are connected to formal sanctions, or informal where the sanctions are also informal. Proper social norms consist of conventions, which apply to a large group such as a whole human society or socio-economic class, and group-specific norms. The sanctions connected to both types of proper social norms are social sanctions, and may include punishment by others and expelling from the group. Aside from these norms, Tuomela also describes personal norms and also potential social norms (these are norms which are normally widely obeyed but which are not in their essence based on “social responsiveness” and which in principle could be personal only). These potential social norms contain among others moral and prudential norms (mnorms and p-norms respectively). The reasons for accepting norms differ as to the kind of norms: • rules are obeyed since they are agreed upon • proper social norms are obeyed since others expect one to obey • moral norms are obeyed because of one’s conscience • prudential norms are obeyed because it is the rational thing to do The motivational power of all types of norms depends on the norm being a subject’s reason for action. In other words, norms need to be “internalized” and “accepted”.

16


2.3.2

Norms in Legal Theory

Within deontic logic, a norm is viewed as an expression of the obligations and rights connected to the role an individual has within a larger social system. This is the second of the definitions taken from the Webster dictionary. The legal theory view on norms corresponds with Tuomela’s r-norms and are backed by formal sanctions. The different schools in legal theory do not differ on the definition of a norm but do differ on the mental dimensions of norms, i.e. on why agents accept and obey norms. In [41] an overview of these different views is presented. The following reasons for norm accepting and obeying are given: • norms are accepted out of fear for the authority issuing the norm • norms are accepted since they are rational • norms are accepted from a sense of duty • norms are accepted since they solve problems of coordination and cooperation I will leave it to the reader to depict these upon the reasons for obeying as developed by Tuomela and described in the previous subsection. 2.3.3

Multiagent Systems Research and Norms

The use of norms in artificial agents is a fairly recent development in multiagent systems research (c.f. e.g., [102], [124], [12]). Even within multiagent systems research different definitions of norms are used. In [39] (pp. 91-92) the following views on norms in multiagent system research are described: • norms as constraints on behavior • norms as ends (or goals) • norms as obligations

Definitions

17

Most research on norms in multiagent systems focuses on norms as constraints on behavior via social laws (c.f. e.g. [23], [78], [102]). These social laws are designed off-line4 and agents are not allowed to deviate from the social laws (except in the work by Briggs, see below). In this sense the social laws are even more strict than the r-norms Tuomela describes which come closest to these social laws. The social laws are designed to avoid problems caused by interacting autonomous selfish agents, thus improving cooperation and coordination by constraining the agents’ action choices. This view on norms is based on the view on norms as developed within game theoretical research such as [115]. In [23], agents may choose less restrictive sets of social laws if they can not find a solution under a set of social laws, thus introducing a possibility for deviation. This approach is close to the approach in [12] where sets of norms are used by an artificial agent decision support system (pronouncer ) to reorder decision trees with the agent having the possibility to refrain from using the reordered decision tree. The reasons behind this are not further developed in [12], in contrast to [23]. However, the title of Briggs and Cook’s article Flexible social laws is deceiving, it is not the laws that are flexible, it is the way they are applied. The laws do not change, it is the agent who decides to apply them or not. The agent is only allowed to deviate from a social law if it cannot act. Thus the authors deny that not acting can be a choice and disconnect the choice of applying a social law from more realistic reasons other than the possibility to act. Work on cognitive grounded norms is conducted in the group around Castelfranchi and Conte (c.f., e.g., [39], [40], [41]) or in research inspired by their work (c.f., e.g., [99]). In [40] norms are seen as indispensable for fully autonomous agents. The capacity for normacceptance is taken to depend upon the ability to recognize norms, 4

In a recent article [103], social laws and conventions are not designed off-line but emerge at runtime. Social conventions limit the agent’s set of choices to exactly one. The agents are not allowed to deviate from the social laws or conventions. Furthermore, a central authority forces agents to comply.

18


normative authorities and on solving conflicts among norms. Since normative authorities are only of importance in the case of r-norms, the agents should also be able to recognize group members to be able to deal with s-norms. In [113] a theory solving conflicts among norms of different categories is developed that can complement the research described in [40]. The origins of norms is not clarified in [40]. However, the possibility of norm deviation is an important addition to multiagent systems research on norms. 2.3.4

Working Definition of Norms

In the current work agents are viewed as having personal norms and coalition norms. The coalition norms are subjective, thus every agent has an individual view on each norm of the coalition. The personal norms emerge from the interaction with the environment. The coalition norms emerge from interaction with the other agents. This synchronization process will result in coalition norms that are shared by all agents and will lie somewhere between the set of personal norms of the individual agents comprising the coalition (this averaging type of norm evolution is commonplace in human subjects [25]). This final state will only be reached if the group does not change and if the individual norms have stabilized. 2.3.5

Learning of Norms

The learning of norms can be divided in two types, viz. the emergence of norms (c.f. e.g. [115] for a description of the emergence of norms from a game theory point of view) and the acceptance of norms. These two types of learning express learning at different levels. The emergence of norms is learning at the level of the social system while the acceptance of norms is learning at the level of the individual agent. Reasons for accepting norms are discussed in the above subsection on Tuomela and the subsection on norms in legal theory. In [40] reasons for the acceptance of norms in multiagent systems are discussed. I am not primarily interested in why agents accept norms. Instead I focus

Definitions

19

on how acceptance of norms changes the decision making behavior of the agents by changing the agent’s definition of the norms of the coalition (norm-spreading) and by the adaption of the agent’s own norms (norm-internalizing).

20


Chapter 3 Internals of an Agent 3.1

Models of Decision Making

In the social sciences, several theories on human decision making have been developed, of both descriptive and prescriptive nature. The most basic model is a model without any deliberation at all, where a single action from an agent is triggered by some characteristics of the state of the environment. The agent has no choice at all but can merely act in the way defined internally (since this means the agent has no autonomy, one might ask oneself if this can be considered to be an agent). This can be equaled to the stimulus-response model that forms the core of behaviorism as developed in [106], among others. A bit more sophisticated is the model of the rational agent used in economics, known as homo economicus. This model has triggered a research field in social sciences known as rational choice theory [38]. The main assumptions of these type of theory are: • individualism: individual agents make their decisions independently • optimality: decisions are based on optimality • self-regard: individuals are concerned with their own well-being only

22


An agent facing a set of alternatives chooses the alternative that will give it maximal gain. An agent therefore needs to have some kind of evaluation function. This evaluation function is utility-based. Application of the evaluation function results in a so-called preference structure. In the basic model, these preferences and utilities are not questioned and are predefined and as such not changeable by the agent itself. Noteworthy is that this model, which under a long period was the dominating view in economic theory, is nowadays under fire in view of evidence from amongst others experimental psychology [1]. More advanced models at the same ontological level include utility-maximizing and utility-maximizing under risk[49]. An even more sophisticated model is the beliefs, desires, and intentions (BDI) model (see, e.g., [21]) or mentalistic model. The terms “beliefs”, “desires” and “intentions” are used here in the spirit of Dennett’s intentional stance [44], i.e., in a folk psychological way, without reference to a specific formalization of BDI models. The beliefs and desires of an agent, together with the context of the agent produce a preference structure. The preference structure suggests the alternative that the agent intends to follow through using means-ends analysis. The intention of the action that follows is to obtain the result of the chosen alternative. This more elaborate model enables us to clarify the choices by looking at the agents’ beliefs and desires, i.e., it gives us a model of the deliberation of an agent. Its main weakness is that it is unclear where beliefs and desires come from. Mentalistic models also have a view on rationality that is somewhat strong. The agent is not only influenced by the complexity of the world, but also limited by its internal complexity. This principle of bounded rationality as developed in [104] limits the evaluation function and the amount of information that can be processed. Furthermore, alternatives are not presented in parallel but rather in a sequential manner. Rather than maximize, agents have to satisfice [105] for they can neither oversee all alternatives, nor evaluate their consequences in a reasonable amount of time. Habits and learned

Internals of an Agent

23

rules (heuristics) are used to simplify the set of alternatives. This does not severely limit the model of rationality in the sense we used it, but rather makes suboptimal choices explainable. What agents can believe and desire is dependent on their style of reasoning [67]. The styles of reasoning have different views on the truth or falsity of propositions as well as the evaluation of possible courses of action. Different situations may ask us to use different styles of reasoning. The idea of an overall rationality and consistency is thus an idée fixe. Different styles of reasoning also arise when there are differences in the experiences of agents. Actions are then not directly linked to beliefs and desires, but instead the style of reasoning or model of deliberation in use forms an important intermediate. These models of deliberation may differ from those used by the observer or in the case of multiagent systems, another agent. The rationality of behavior is no longer the model of deliberation that is defined as rational in the above mentioned sense, but rather if the specific model of deliberation used by the agent under consideration is employed in a correct sense. Irrational behavior is behavior that violates the model of deliberation the agent is using for the situation at hand. More strongly put, irrational behavior can only be displayed by agents that are rational, i.e., in general employ a model of deliberation in the specific situation.

3.2

Considering Other Agents

The models presented thus far are all based on what Weber [127] calls purposive rationality (“Zweckrationalität”). It is rationality on a cognitive instrumental level; only the goals of the agent itself are considered in the pursuit of self-maintenance. Decision making at a social level is impossible without considering other agents. In [100] a distinction is made between theories of rationality (equivalent to teleological action models) and theories of sociality that explain why people sometimes act as they do because of their aware-

24


ness of certain others. Theories of rationality can be based on a BDI model of the deliberation of an agent. Rationality in this context means that an agent’s beliefs, desires, and intentions are coherent and that its actions, plans and goals are coherent.1 Cooperation is defined by Schick as an action chosen in view of the future, in consideration of other agents’ possible choices. Therefore an agent has to have a model of the other agents at least at the BDI level. The reasons for taking others into consideration may differ. From a cognitive rationalist point of view, they can be linked to self-interest, e.g., an agent may cooperate because it believes it will pay off in the long run, or because it is afraid other agents will punish it for not cooperating. Choosing socially is another matter. Choosing socially is defined by Schick as: An individual considers the goals of others before considering its own goals.2 Considering other agents’ goals does not mean an agent adopts them as its own goals. The other agents are modeled neither as inanimate objects in the environment, nor as objects that cause changes to the environment, but rather as objects that change the environment according to their own goals. Commitment does not enter Schick’s ideas until norms and values are considered. A promise or contract in its own right is not enough for a commitment. Rather, they are seen to activate a norm or ideal an individual has. It is the individual’s own “ethics” that makes it committed. This ethics can be seen as a second-order desire, since it focuses the individual on 1

The concept of rationality is problematic, but I do not wish to delve deep into this matter here and instead refer the reader to proceedings such as [52] and [94], or monographs such as [61], [65], and [10]. Two well-known definitions of rationality also mentioned by Schick are the definition by Plato (to be rational is to act upon knowledge only, on nothing but an understanding of what the world is like) and Kant (as taken by [10] to state that rationality means having a reason for behavior). 2 This is a very strong definition of choosing socially. Later I will introduce the work of Margaret Gilbert [59], [60], [61] that describes various kinds of sociality lying between the extremes of pure individualism (or egoism, i.e., the cognitive rationalist tradition) and total altruism (as used by Schick in the above mentioned definition).

Internals of an Agent

25

what it wants to want. Control at this level is the highest level of autonomy an agent can have and this enables as mentioned before the highest level of self-organization a society of agents might reach. I will elaborate on the concept of “want to want” (see also [83]) in the section on levels of agenthood.

3.3

Modeling Other Agents

An ambitious example of an ontology describing the levels on which an agent may model other agents is developed by Baressi and Moore [7]. They outline a hierarchy consisting of four levels of representations of intentional relations (activities involving an agent, an object, and a directed activity). The authors use this hierarchy to separate animal cognition from human cognition and also to offer an explanation of autism in these terms. At the first level, first person information on the intentional relations of the agent to an object is treated differently from third person information on intentional relations of others. An agent therefore has two types of knowledge about intentional relations. At the second level, these two types of information may be integrated when a particular intentional relation to an object is shared with another agent. The similarity between the agent and the other agent with whom it shares the intentional relationship need not be recognized by the agent. The agent will act differently depending on whether it shares the intentional relation or not. At level three, both current information of an intentional relation and imagined information (consisting of recollections of similar previous situations) is input to the interpretation process. Information input instead of resulting in an action results in a mental image of the intentional relation. The self/other distinction becomes explicit and the agent can be said to have a model of the other as an intentional agent. Both types of information are still linked to the current situation.

26


The fourth and final level is a level where both inputs into the interpretation process come from imagination. There is no current first or third person information. The intentional relation is abstractly represented in an orientation toward a mental object. The agent now has a model of the other agent as a mental agent. Although the authors point out that this level is not the final level (e.g., reflective agents are another level), for the purposes of their article this is not further worked out.

Chapter 4 Sociological Theories 4.1

Habermas’ Theory of Communicative Action

In his monumental work on the theory of communicative action [65], Habermas discusses and in a sense unifies several different sociological schools of thought (i.e., Weber, D¨ urkheim, Mead and Parsons) using and criticizing among others Toulmin’s argumentation theory [112] and speech act theory [3], [101]. Habermas distinguishes four action models. Each action model makes presumptions about the kind of world the agents live in which has consequences for the possible modes of rational action in that model. The first action model is teleological action. Agents are goal directed and try to maximize their choice of means to obtain a goal. This is the rational choice model. The central issue in this action model is the choices the agent makes between different action alternatives, based on maximizing utility. Agents can thus - try to influence the world, and the rationality of the behavior of the agents can be evaluated with respect to the efficiency of their behavior. Adding other agents, with respect to whom the agent acts in a strategic manner (strategical action model), to the decision making model does not change the ontological principles. The agents may need to model the desires and actions of the other agents but these are still part of the objective world of existing states of affairs. Agents act with respect to this world according to their beliefs about the existing states of

28


affairs and their intentions to bring about desired states of affairs in that world. The second action model is the normatively regulated action model. Social agents are assumed to belong to a group and follow the norms that are obliged to be followed by members of that group. Following norms is taken as to behave according to expectations. The objective reality is extended by a social reality of obliging norms (acknowledged as such by the group). The rationality of the behavior of an agent is not only related to the objective reality (teleological and strategical action model), but also to the social reality. The conformity between the norms of the group and the behavior of the agents and the relation between the norms and the generalized interests of the agents (and thus if it is wise of the agents to confirm to those norms) are part of this social rationality. Agents act with respect to an objective world and a social world, namely the normative context that defines the possible interactions and legitimate interagent relationships between the agents. The third action model is the dramaturgical action model. In this action model the inner world of the agents is considered. Based on Goffman’s [62] dramaturgical analysis of social life, this action model has as a core the presentation of the self of an agent to an audience. This representation of the self may or may not be truthful. The agent makes use of the fact that its inner self is only admissible to itself. The inner self is defined as the constellation of beliefs, desires, intentions, feelings, and needs of an agent. Habermas views this inner self as a reality in its own right. When presented in a truthful and authentic way, and at the same time connected to the shared evaluation criteria and interpretations of needs, the subjective point of view of the agent can gain an intersubjective value. Truthful is not the same as true in objective sense, opening the door for lying and manipulation or insincerity. Agents act with respect to an objective world and a subjective world formed by the totality of subjective experience to which the agent has a privileged access. Examples of application of this action

Sociological Theories

29

model in the field of MAS include lying agents and believable agents (see 7.3 below). The fourth and final action model is the communicative action model. This action model unites the three functions of language specified by the three previous action models. In the strategical action model, language is used by an agent to reach its own goals possibly via influencing other agents by use of language, the normative action model uses language to actualize already existing normative agreements and the dramaturgical model uses language to allow for one to express oneself. In the communicative action model, language is used to bring about mutual understanding on all three previous levels. The agents use language to claim the truth of their utterances, the normative correctness of their speech acts in view of the context, and the sincerity of their intentions being formulated. Testing for rationality of actions is here no longer the privilege of the observer, but is done by the agents themselves to realize a common definition of the situation described in terms of relations between the speech act and the three worlds (i.e., the objective, social, and subjective world) this speech act has relations with. In the cooperative process of interpretation, all participating agents have to incorporate their own interpretation with that of the other agents so that the agents have a sufficiently shared view of the external (i.e., objective and social) world in order to coordinate their actions while pursuing their own goals.

4.2

The Micro-Macro Problem

The micro-macro link problem is the problem of the relation between the behavior of individuals (micro) forming a group and the behavior of the group as a whole (macro). Where cognitive science places the individual at the center, social theory has two main streams of research. One places the individual at the center of attention, reducing all behavior of collections of individuals to the behavior of the individuals themselves (methodological individualism or reductionism).

30


The other places the collective behavior at the center of attention and in fact makes the behavior of individuals dependent upon the group to which they belong (macro sociology or holism). In the following we will look at two approaches toward the problem of the relation between the behavior of (artificial) agents and that of the multiagent system. Cognitive science models are usually derived from models within the methodological solipsistic and functionalistic tradition, causing agents to be viewed as self-enclosed modules, isolated and strictly autonomous from the environment in which they function [76]. Working on coordination starting from a cognitive theory therefore means that the problem is tackled from an individual perspective. Individual rationality is used to explain why agents behave “socially”. Agents are also viewed as planning agents [21] who are goal directed [98]. The influence of the information processing view [86] can be felt here. The model of communication is based on speech act theory [3], [101]. In speech act theory, the agent is solipsistic as well, trying to change the world in order to achieve its own goals. The study of action in cognitive science is a study of problem solving, how an individual agent can construct a sequence of operators (i.e., a plan) in order to change the environment in such a way that a goal is achieved. Communication is viewed as a planned action sequence, which is not meant to change the (physical) environment, but rather the beliefs and intentions of other agents. Social activities are explained by assuming mutual beliefs, intentions, and knowledge which guarantee that a set of individual agents behaves on the macro level as designed. Assuming mutuality on the level of knowledge and intentions implies that the agents are able to analyze each other’s behavior. The coordination problem can thus be restated as the problem of how to reach (a sufficient level of) mutuality among individual agents. Mutuality is not reached through a process motivated by individual intentions or values, but rather by hardwired intentions such as benevolence assumptions [48] or mutuality that is assured beforehand

Sociological Theories

31

(e.g., [70]). Rational choice theory and game theory, which both have gained some following in sociology, are frameworks where the behavior of the individuals are defined while the behavior of the social system as a whole is a consequence of the interaction of the individuals and remains undefined. In order to steer the behavior of the social system, the behavior rules of the individuals have to be changed. Most sociologists start from an opposite view on the relation between individuals and the social system. They view an agent’s behavior as defined by the agent’s position within the social system. One example is Marx’s theory of classes [80]. Individuals are either capitalists or employees and the position of the individual determines his actions (e.g., maintaining the current situation or working toward a revolution). The influence of the system on the behavior of the individuals is mediated by the norms and values which members of a social system necessarily share. Coordination takes place on the system level and is imposed on the individuals who have no say in this. Giddens [58] lays out the foundations for an intermediate theory. In his theory of structuration the relation between the system and the actions of the individuals of the system is viewed as circular. The system is embedded in and structures the actions, which in turn reproduce the structuring properties of the system. Conte and Castelfranchi [39] agree with Giddens’ theory of structuration but show that Giddens does not specify how the two-way link is achieved. Conte and Castelfranchi propose to use cognition as a medium between the micro and the macro. However, they fail to take the function of norms beyond that of role descriptions of individual members. Although they agree that the normative dimension is crucial to any theory of social action, they do not link this dimension to the cognitive processes of the individual as a separate level of deliberation.1 Since norms and values are harder to change than beliefs, desires, and intentions and 1

In more recent work (e.g., [40]), the authors change their focus to this aspect of normative agents.

32


therefore influence behavior over a longer time-period, a level separate of the BDI level is needed [84]. Looking at individuals as being value-realizing systems makes understanding what is going on in the real social world more feasible [79], so this approach could very well help us build human-like agents in an artificial social world that is more like the real world. And as Habermas stated in his theory of modernity [65], reasoning about values and norms is not linked to the instrumental rationality used on the goal level. The legitimacy of values and norms is questioned, not the optimal path to some goal. If we want to use artificial agents to model human behavior or behave with a self-organizing capability comparable to that of humans, we need a model of deliberation in which norms play a role, separate from the intentional level (BDI and goals), the planning level, and the reactive level. This model will be developed below.

Chapter 5 Levels of Agenthood In this chapter, I will first describe various approaches to categorization of agents with respect to their reasoning capabilities. Dennett [44] has developed a ladder of personhood. Although this ladder was created with a goal different from ours, some of Dennett’s ideas can be used to create a typology of agents. I will compare this to the work by Carley and Newell [33], the levels of agents that Conte and Castelfranchi [39] use, and the work of Werner [128]. After this, I introduce my typology of agents and discuss the similarities and differences between my model and the previously described ones.

5.1

Dennett’s Ladder of Personhood

The first step of Dennett’s ladder of personhood is formed by three different, but interrelated concepts, namely rationality, intentionality, and stance. An intentional system is a system that can be analyzed using the intentional stance. The intentional stance assumes that a system (or agent) behaves rationally and intentionally, i.e., as if it has beliefs, desires, and intentions. Dennett also links communication to this level of personhood. Communication, according to Dennett, is founded on the presumption of the rationality of the agents we communicate with. One step up the ladder is the level of reciprocity. An agent on this level is able to reciprocate the stance, in that it is able to view other agents as intentional systems. The agent must therefore

34


be able to form beliefs, desires, and intentions about beliefs, desires, and intentions and becomes a second-order intentional system. The third step of the ladder adds the concept of verbal communication to the capacities of the agent. It introduces a third-order intention, an agent has the intention that another agent, when receiving the first agent’s message, recognizes what the first agent would like the receiving agent to do. The example Dennett presents is that of a conductor on a bus ringing the bell on the bus three times to notify the driver that the bus is full and should not stop at the next bus stop. This third-order intention results in what Dennett calls “genuine reciprocity”. The fourth and final step of Dennett’s ladder of personhood is consciousness, which Dennett equates with Frankfurt’s [54] notion of reflective self-evaluation, which results in second-order desires. This is an idea used by Miceli and Castelfranchi [83] who use the expression “want to want” to build a formal model of normative expressions. And as mentioned earlier, Schick [100] equates an agent’s will (i.e., an agent’s ethics; its norms and values) with second-order desires. Dennett more precisely calls this top step of his ladder of personhood “moral personhood”.

5.2

Carley and Newell’s Model Social Agent

Carley and Newell [33] analyze what is needed to build an artificial social agent. They distinguish two dimensions along which agents can be categorized: the information processing capabilities of the agents and the knowledge the agent possesses. These dimensions are also used to categorize theories on the behavior and decision making of artificial and human agents and simulation models on organizational decision making and multiagent research projects (see also [31]). Figure 5.1 displays a matrix based these two dimensions.1 The most simple and least human-like agent one can imagine using these dimensions is an agent who is omnipotent (i.e., without any 1

The matrix is adapted from [33].

Levels of Agenthood

35 Knowledge

Cognitive architecture

increasingly limited capabilities

increasingly rich situation real interaction

social social structural goals

cultural histrorical

nonsocial task

multiple agents

omnipotent agent

goal directed

rational agent

reasoning information gathering

class organizational historically timing exchange situated differences goals constraints theory turn taking modelling of others emergent social learning from miscommunication social norms mobility cognition others competition negotiation

boundedly rational agent

satisficing planning adaptation

cognitive agent

compulsiveness group think

emotional cognitive agent

habituation variable performance

group making social planning coercion

protesting courting

social interaction

play mob action

altruism information networks

moral obligation

group automatic conflict responses to status cues campaining team player

role emergence

develop language institutions norm maintenance ritual maintenance

Figure 5.1: The agent categorization matrix limitations at all on its information processing capabilities) possessing only knowledge about the task it is working on. At the other extreme, we find an agent who is both emotional and cognitive, with knowledge of the task, as well as other agents, interacting in real-time within a social structure with social goals and cultural and historical knowledge of that social structure. It is this agent that Carley and Newell label model social agent. Between the omnipotent agent and the model social agent all sorts of agents can be distinguished, which in turn can be linked to different types of social theories. There is also a tradeoff between information processing capabilities and behavior: the more severe the limitations on the information processing capabilities are, the more complex behavior may emerge. For example, the omnipotent agent does not have to gather information, since everything is already known to the agent. As for the knowledge dimension, the more knowledge an agent has, the more detailed its mental model of the world can be and the more realistic it will be, as compared to the richness of the mental models of humans. With a richer men-

36


tal model the agent has a richer repertoire of actions to choose from and its set of goals can be more complex. For instance, if an agent has no knowledge of social structure, the agent’s place in the social environment does not have to (and even cannot) be determined. Now let us turn to a detailed look at both dimensions. As already mentioned, the omnipotent agent is the starting point of the information processing capabilities dimension (starting from the least limited site). This type of agent knows all there is to know about a situation, will take the actions necessary to obtain its goal, and is able to change its environment. This type of agent is equivalent to the homo economicus. When we see an agent as not omnipotent with respect to its task environment, but rather as an agent with a body of knowledge of its own, with respect to which it takes the most appropriate action to obtain its goal, we have an agent at the knowledge level [85], i.e., a rational agent (in the teleological action model sense of the word). The agent can alter the environment, but it can also interact with the environment to acquire knowledge. When the agent’s attention is limited, and therefore cannot process all available knowledge, we have a bounded rational agent. Within the limits of its abilities and knowledge, the agent tries to obtain its goals. It has an internal model of the task(s) it is working on, and has a cognitive structure consisting of a long-term memory for the permanent knowledge and a short-term memory for knowledge of the current situation. Memory units, called chunks, are linked in an associative hierarchical way. This hierarchy must be traversed when working on a task. The task can be split into problems and subproblems. Thinking includes both problem solving and learning, and takes time. The cognitive agent is a bounded rational agent with a fully specified cognitive architecture (the Soar architecture [74] is an example of such an architecture). A cognitive architecture is a set of information processing mechanisms used to generate all behavior. The mechanisms embody the processing limits. The limits are not as general as the limits of the boundedly rational agent, but are operationalized. The emotional agent is the

Levels of Agenthood

37

most limited agent in this framework. Emotions influence cognitive decisions. For example, if we do not like a person, we will prefer to work together with another person rather than the one we do not like, or even prefer to fail at a task (which cannot be predicted when taking the cognitive agent as a starting point). Unfortunately, the connection between rational and cognitive agents and emotion-based theories has traditionally been very weak [87]. However, from the mid-eighties onwards, there have been more research efforts as is witnessed by the appearance of a journal (called Cognition & Emotion) in 1986, and [89]. The last couple of years, various workshops have been organised (cf., e.g., [26], [27]). The other dimension is the knowledge dimension, or more accurately the type of information on the situation the agent is in that is used in the decision making behavior of the agent. The authors distinguish six situations. The agent can make use of cultural-historical knowledge about the situation it is in. The situation is affected by and a result of a historical process leading to a specific culture, embodied in norms shared among the members of a group. If the agent does not take the cultural-historical knowledge into consideration, it can still be aware of social goals. Apart from the goals related to the task at hand and the goals related to self-maintenance and enhancement, the agent also has goals that have to do with social maintenance. These social goals can be irreconcilable with the goals of the individual or even of the task, thus influencing the problem solving behavior of the agent. Lifting the constraints of the social goals leaves us with constraints due to the social structure the agent is in. The agent has knowledge of the structure of the group it is in, and the position of the agents it interacts with. If the agent has no knowledge of the social structure it may be engaged in face-to-face interaction. The agents do interact, but the interactions are not directly with individual agents but mediated via the environment. The interactive situation limits the time an agent has. Absence of such a time limit results in an agent in a multiple agent situation. There are other

38


agents in the environment, with their own goals and capabilities and displaying goal-directed behavior. The last situation is the nonsocial task situation. The agent only has knowledge of the task and the task environment. Other agents may exist in that environment but are treated just like inanimate objects.

5.3

Conte and Castelfranchi’s Agent Typology

Conte and Castelfranchi [39] describe four types of agents, namely reactive, plan autonomous, goal autonomous, and social autonomous agents. Conte and Castelfranchi define an agent as a goal-directed system that performs actions based on its perception of the world and applying its internal representation of causal relations between actions and their consequences to achieve a given intention. Conte and Castelfranchi claim that a reactive agent is not very profitable in a multiagent context since they are not socially responsive. Even stronger, a reactive agent cannot be an agent according to Conte and Castelfranchi because of their definition of an agent, since a reactive system links actions to external stimuli without any internal processing. They more appropriately use the term goal-oriented systems to refer to reactive agents. Goal-autonomous agents are autonomous with respect to their goals and ends, whereas plan-autonomous agents are autonomous with respect to their actions and plans. A planautonomous agent does have goals, but its goals are dependent upon others. The agent plans to achieve a goal others want it to achieve. It is unclear whether Conte and Castelfranchi mean that other agents or the designer determine whether an agent is to pursue a goal. The social autonomous agent is considered as a special case of the goal autonomous agent without the benevolence assumption [48], so that this type of agent can autonomously decide whether to reward requests related to goals of other agents.

Levels of Agenthood

5.4

39

Werner’s Agent Typology

Werner [128] also sketches different types of agents when developing a general architecture of multiagent systems. His fundamental assumption is that coordination and cooperation in the true sense of the word is only possible by way of exchange of information. Werner defines an agent as an entity that is guided by some strategy that controls its behavior. An agent has three types of representational states; an information state (containing the information the agent has about the world), a strategic state (containing the agent’s control information, i.e., information that guide its actions), and an evaluative state (containing the agent’s evaluations and guiding its reasoning). The content of these three states may differ and these variations are used to describe various agent types. The most basic type of agent described by Werner is the solipsistic agent. This is an isolated agent that is truly autonomous (in the sense that Castelfranchi used in [35] and can be seen as equal to the omnipotent agent as described by Carley and Newell [33]) without interaction and coordinated behavior. The agent accepts no input and therefore interacts neither with other agents, nor the environment. Note that the use of the term solipsistic as defined by Werner is more strict than its use in other parts of this thesis. An agent whose strategic state contains solely rules that directly couple environmental changes to agent actions is called a purely reactive agent. Other agents are treated as mere objects. If the reactive strategies of the agents have co-evolved or are preprogrammed so as to achieve some social goal this may result in coordinated behavior. The environment is in a sense programming the agents and the agents also program each other and themselves by changing their common environment. The kinds of coordination that these agents can achieve is extremely limited. Referential agents are agents that can communicate parts of their informational state to other agents and receive informational state information from other agents. These communicated symbols are in-

40


terpreted by the receiving agent, whose information state may be changed as a consequence. Intentional agents are agents that can not only communicate parts of their informational state, but also parts of their strategic states (i.e., their intentions). These communications may result in changes in other agents’ strategic states. This makes it possible to directly influence, control, and coordinate the agent’s actions (by slightly stretching its meaning, one might say that this is what Habermas calls the dramaturgical action model). Social agents have not only intentions that guide their actions but also the ability to take on various roles in social settings (for application of these principles in agent-based systems cf., e.g., [72]). These roles involve both state information, strategic information, and evaluative information. The adoption of a role then forms and constrains the agent’s overall representational state. To be able to adopt a role the agent must have all abilities required by the role.

5.5

Levels of Sociality

The discussion about what is to count as social phenomena has kept sociologists busy from the very start of their profession’s existence. In this section the work of Gilbert [59], [60], [61] will be the main source for our account of the debate on sociality. I will give the conclusions of Gilbert and not the arguments supporting the soundness of her intuitions about the notion of sociality. Instead of striving for an all or none definition of sociality, Gilbert proposes the idea of degrees of sociality. Her approach is reminiscent of the earlier described matrix proposed in [33]. Sociality is defined by Gilbert as a causal or mental connectedness between two or more agents. The degree of sociality of a phenomenon is dependent on the degree of mental or causal connectedness between the agents involved. The distinction between causal and mental connectedness gives us a first criterion for distinguishing degrees of so-

Levels of Agenthood

41

ciality. Causal connectedness refers to situations where there is no interaction based on common knowledge and/or mutual expectations. Interactions between agents occur through the changes their individual actions cause in the shared environment. Mental connectedness refers to situations where agents have some kind of model of other agents. The nature of the interaction between agents depends only on the notion of singular agency. An agent has its own beliefs, attitudes, goals, and intentions, some of which it may share with others, in a weak sense. Sharing in the weak sense means it may be common knowledge among the involved agents that all of the agents have beliefs, attitudes, goals, and intentions. It does not license the agents to interfere or to involve themselves in each other’s progress in achieving the goal. Of course each partner-in-action could be motivated by altruism (an example of this in the realm of multiagent systems is the previously mentioned benevolence assumption) to help each of the others reach their own individual goals and thereby reach its own goal. But according to Gilbert these motives do not properly capture the flavor of performing an action together in the strong sense. These motivations are essentially asocial. Agents are driven by their own personal beliefs, desires, and intentions. Some of these may be quite other-directed but the agents are essentially still self-centered. To show the essence of sociality, Gilbert introduces the concept of plural subjecthood. Plural subjecthood is associated with joint action. Let us first see what Gilbert means by joint action or two or more agents sharing in an action. Joint action is not properly described as doing the same type of action independently. What kind of dependency is necessary to speak of joint action is the next issue to address. A possible candidate is spatio-temporal contingency. An account could be given in behavioral terms such that each constituent of the plural subject acts in a particular way, i.e., in pursuit of the common goal, while trying to do the acts together with the other constituents. Two objections could be given to this.

42


Firstly, acting jointly does not entail that all partners-in-action count as doing the action when considered in isolation. Second, the goal of a plural subject is not to be equated with a goal that each of its constituents has adopted as its own goal. All that can be said is that each agent must intentionally act in its capacity as a constituent of a plural subject of a certain goal. Plural subjecthood is formed by the mechanism of joint commitment. Joint commitment is reached by setting up some conditions required for doing something together by each party making certain things clear to the other party. It is not enough that they are both aware of each other’s willingness to share in action: they both have to express it openly to each other in a situation of common knowledge. Doing something together (in the strong sense) involves rights and duties that are something other than moral rights and duties. For example as a constituent of a plural subject it is proper to solicit information on, as well as to interfere or involve oneself in, the progress of the other’s contributing to the achievement of the joint goal. Gilbert uses “walking together” as an example to illustrate these concepts. Suppose a person is walking all by herself. If another person were to start walking alongside her, this does not make them walking together, just walking alongside of each other. Even if both of them have as a goal to be walking alongside each other, this does not change. Gilbert labels this situation as weak shared personal goals. Making their personal goals common knowledge (i.e., both of the persons involved is aware of the goal the other person has) changes the situation to strong shared personal goals. The goal each person has is still at the individual level, even though the goals the individuals have are the same. Both parties may agree to be part of a plural subject (as opposed to a single individual as was the case previously) with the goal of walking together. A plural subject consists of at least two persons, pursuing the same goal as a consequence of being part of the plural subject of that goal. This implies that all parts of the plural subjecthood have an obligation to pursue the goal, entitling

Levels of Agenthood

43

the other parts to expect a certain performance. So, if two persons agree on walking together and one person starts to draw ahead, the other person may point out this so as to obtain a walking together situation again. Entering a plural subjecthood is done via a conditional commitment, so that a person is only committed if and only if the other persons have expressed the same conditional commitment.

44


Chapter 6 Proposed Agent Typology The hierarchy of types of agents proposed here is based on different reasoning capabilities. The exact makeup of this reasoning capability is not important to the present purposes. I will link the reasoning capabilities to the level of autonomy of an agent. In the following sections, I will link the types of agents to the possible content of communication and kind of models that can be used to model other agents. Combining the reasoning capabilities with reflexive monitoring will result in learning on various levels. Agent types: Type I

Type II

1

reactive agents perception is directly coupled to action: no deliberation, only stimulus-response associating. A reactive agent has no means of influencing the environment in a preconceived way. The autonomy resides completely in the combination of environmental cues and properties of the system. plan autonomous agents1 the agent’s activities are oriented toward a goal state that cannot be changed by the agent itself. The agent chooses actions from a repertoire of actions by means-ends rea-

Or rather agents autonomous with respect to the creation of plans. The term plan autonomous agents is taken from [39].

46


Type III

Type IV

2

soning. The agent is triggered by a request from another agent to achieve a state of the world or by the state of the environment, and is independently motivated to pursue its goal whenever it receives a request from another agent. Goals and interests are not distinguished. Whenever an agent is requested to achieve a goal, that goal becomes an interest. Plan autonomous agents have more autonomy than reactive agents in the sense that they may choose how to achieve a certain state of the world. They still do not have the notion of goal. They are equipped to achieve a given objective and have a repertoire of actions that they can aggregate into a sequence (i.e., a plan) that— in the particular circumstances in which they receive the request—will achieve the objective. goal autonomous agents2 goals are not just triggered by requests (thereby becoming interests) but must also be linked to a goal the agent already has. A goal autonomous system has the autonomy to determine what its “prevailing interest” is, considering its goals. It will judge which states of the world are its interests by evaluating what it provides in terms of degree of goal satisfaction and goal priority. I will call the reasoning capability of these agents strategic reasoning. norm autonomous agents these agents choose which goals are legitimate to pursue, based on a given system of norms. The agent has the autonomy of generating its own goals and to choose which it is going to pursue. Besides, the agent is equipped to judge the legitimacy of its own and other agents’ goals. When a goal conflict arises (not to be confused with interest conflict), the agent may change its norm system

Or rather agents autonomous with respect to the setting of goals. The term goal autonomous agents is taken from [39].

Proposed Agent Typology

47

thereby changing priorities of goals, abandoning a goal, changing a goal, generating another goal, etc . I will call the reasoning capability of these agents normative reasoning. Norm autonomous agents generate norms they can use to evaluate states of the world in terms of whether or not they could be legitimate interests. Legitimacy is a social notion and is in the end determined by the norms of the agent with respect to the agent society it is a part of. All types of agents also possess the capabilities of the types preceding it as illustrated in figure 6.1. The typology of agents I have intro-

norm autonomy goal autonomy plan autonomy reactivity

Figure 6.1: The proposed agent typology duced can be linked to the above described ideas of Dennett, Conte and Castelfranchi, and Werner. Agenttype I can be seen as equivalent to Conte and Castelfranchi’s reactive agents, and Werner’s purely reactive agents. Types II and III share not only their name with Conte and Castelfranchi’s plan autonomous and goal autonomous systems,

48


they are equivalent to these systems. Type IV is comparable to Dennett’s moral personhood. A need for models on this level is signaled by several researchers (cf., e.g., [53], [68], [33], [128], [70], [64]) who have noted the limitations of Newell’s knowledge level [85]. The knowledge level can be used to model systems up to goal autonomous systems, but for normative evaluations, another level may be needed. This level received different names, such as organizational level [53], cooperation knowledge level [68], and social level [70]. However, since all these researchers start from an individualistic approach, they do not take their ideas to the (sociological) limit. Indeed, Werner [128] and Carley and Newell [33] provide the most substantial attempts at modeling truly social agents. The idea that the behavior rules for agents to behave in a “social” way are not located at the knowledge level but at a separate level implies that those behavior rules have a different ontological standing. However, the original description of the knowledge level is perfectly at ease with different types of knowledge. It is the formulation of the rationality principle that is the problem, since this is based on the solipsistic paradigm, allowing only for a goal-directed agent. The system modeled at the knowledge level is indeed supposed to be a unitary system, and therefore in Newell’s [85] analysis of systems in the social band, these social systems are built from several agents but are analyzed as being just one meta-agent. Another way around the problem with the knowledge level would be to relax Newell’s demands and allow for both multiagent situations and several definitions of rationality. The original knowledge level analysis would be the special case with only one agent and (thus) rationality being equal to efficient goal-directed behavior. Social knowledge is then just one of the many forms of knowledge. This approach is used in [64] to develop a society level view of multiagent systems. In the typology proposed above, the model an agent may develop of other agents is implied by its agent type. Agents are able to model the decision making model of other agents as up to and including

Proposed Agent Typology

49

the agent’s own level in the typology of agenthood (this principle is also used in [125]). Reactive agents form an exception to this. These agents can view other agents only as mere objects. Reactive agents act on observations of the environment (or even more strongly, impulses from the environment), no matter how the state of the environment has come about. Its model of deliberation is the stimulus response model. Plan autonomous agents may see the other agents as planning objects (or reactive agents since the capabilities are inherited from the previous level). The behavior of acting objects may change the state of the environment. The agent’s own model of deliberation is means-ends reasoning. Having a model of the model of deliberation of other agents makes it possible to make use of this when in need of an action or plan one cannot perform oneself. The other agent then becomes a means to reach an end. The plan autonomous agent can model other agents driven by means-ends reasoning or responses to stimuli. In the first case, the agent may send a request to the other agent, in the second case it can only manipulate its environment. Goal autonomous agents can view other agents as at most agents possessing beliefs, desires, and intentions. The model of deliberation is a BDI-like model extended with norms (norms being constraints that are not subject to reflection). The agent can, in addition to sending a request to another agent or manipulating its environment, steer the choice behavior of another agent at the level of goals by changing the other agent’s beliefs. This can be done for example by using power or argumentation. Norm autonomous agents are able to model other agents as agents with norms of their own. The model of deliberation is BDI-like, extended with norms (norms in the strong sense as discussed above and subject to reflection) and the need to share these norms with others. The agent makes an appeal to the norms of the other agents. However, this appeal is not driven by its own interests. Having motivations beyond one’s individual interests is having motives that are

50


supra-individual and therefore social. This is the level of Habermas’ communicative action or Gilbert’s plural subjecthood. Entering a plural subjecthood naturally presupposes the existence of other agents with the same urge.

Chapter 7 Multiagent Systems Research In this chapter I will describe various multiagent systems research schools, ranging from reactive agents to artificial social agents. In each section I will state which type of agent is used in terms of my agent typology and which questions stay out of reach of the researchers because of this choice.

7.1

Reactive Agents

Research based on reactive agents (agents of type I in the proposed agent typology) is widespread. The paradigm of reactive agents is closely tied to research on robotics and artificial life, cf., e.g., the work of Brooks [24] and Steels [108]. It is not limited to these domains, however. Other applications include for example simulations of animal behavior and simple assistants currently replacing simple help functions. Proponents of using reactive agents state some advantages these agents have over cognitive (or intentional or deliberative) agents. The absence of a cognitive apparatus renders an explicit model of the world obsolete. They also lack an idea of sequentiality and are thus unable to build plans. Their simplicity makes them dependent upon each other seen from the level of the behavior of the system as a whole. Reactive agents usually operate in large groups where more than one agent can react in the same way to the environment, thus making the multiagent system robust to any loss of agents. Reactive

52


agents are especially useful in worlds where the rate of change is faster than agents can update their world model. The absence of a world model (including a model of the other agents) renders this type of agent unsuitable for research on emerging social structures since the structure observed is dictated by physical properties of the environment (including the distribution of the agents over the environment) and has no meaning to the subjects. The structure is thus only in the eye of the observer, as is the case in the teleological and strategic action model described in the section above on Habermas. The same holds for the rationality of the agents.

7.2

Intentional Agents

The intentional type of agents (or agents of type III in my agent typology) is in many respects the opposite of the reactive type of agents. Intentional agents have an explicit model of their intentions (what the agent wants to achieve in the current situation), desires (what the agent would like to achieve in general) and beliefs (a model of the environment, which may or may not include models of other agents and the agent itself). BDI agents form a subset of this type of agents. These agents act following the strategic action model. As a consequence of this, these agents are not to be considered social agents (in the norm autonomous or normative action model sense), even though an agent architecture like GRATE* [69] makes use of social concepts such as joint intentions and joint commitments. One could say that these concepts as used by Jennings are related to the dramaturgical action model.

7.3

Believable Agents

Probably the first thoughts on believable computer agents is the Turing test (although the original paper [114] was on distinguishing female from male humans by analyzing discourse via teletype terminals,

Multiagent Systems Research

53

it was later generalized to an operationalization of a test for intelligent behavior). The idea that a human cannot distinguish between the reactions of another human and those of an artifact is the core of this test. Unfortunately, the test focuses on person-to-person interaction, not on interactions in a social setting. In the development of artificial intelligence as a scientific discipline, this idea was consequently narrowed down to intellectual challenges and solving problems. The research on believable agents focuses on creating interactive agents that give users the illusion of being human. Application domains include among others human-computer interaction, interactive entertainment, and education. Believability is accomplished by convincing the humans interacting with the agents to build psychological models of the agents by observing their behavior. The building of these models is enhanced by having the agents express their emotions in their behavior and equip the agents with clearly distinguishable personalities. This has consequences for the agent’s internal model of deliberation. It has to have knowledge of both emotions and know how these can be expressed. This research is connected to Habermas’ dramaturgical action model. If social behavior is of importance in the domain in which the agents act, the agents also have to have knowledge of norms and how these may be expressed. This means the agents have to be of type IV. Examples of this type of research include the Oz project at CMU [9], the Agents group at MIT [77], the Virtual Theatre project at Stanford [66], the work by Rizzo et.al. [96] and commercial activities such as Virtual Personalities.1 The research projects mentioned do have an account of the role of emotions and social norms in agent behavior, but do not have agents that are norm autonomous in the sense as described in the agent typology proposed here. These agents typically fall within the dramaturgical action model. 1

See www.vperson.com.

54

7.4


Artificial Social Agents

Jennings and his fellow researchers have been working on the balance between individual autonomy of agents and optimal behavior of the group they form. The focus of their work changed from commitments to social rationality. The main idea is the same, how to have a collection of autonomous agents behave as a reliable and efficient group. Since rationality is defined in the teleological spirit, this research uses agents of type III of my typology. The research group around Castelfranchi works on dependence relationships and commitments as mechanisms for coordination [36]. They argue against game theory and comparable models for social interaction and argue for more influence from rivaling social theories that have a richer conception of social interaction. Put in the terms used in the section on Habermas, Castelfranchi’s group points to the importance of the dramaturgical action model (which is by definition not analyzed in the game theoretical tradition) and the normative action model.

Part II Norms, Adjustable Autonomy and Simulations of the Spreading of Norms

Chapter 8 Norms Can Replace Plans Abstract We take the position that sophisticated planning may often be replaced by norms in complex systems that include agents with adjustable autonomy. This stance is substantiated in two steps. First, some observations of what autonomy can be and mean to such systems. Second, by a brief exposé of norms and their implementation in systems for artificial decision making.

8.1

Introduction

Complex dynamic environments pose important problems to developers of automated and situated systems. Since it is impossible at the time of development to predict all possible states of the environment, such systems are inevitably non-optimal. One solution to this problem is to limit the autonomy of the system at the outset, and accept that the system has to cooperate with human supervisors. This yields a so-called social space, or artificial ecosystem, in which humans and artificial agents depend on each other to some extent. If the level of autonomy of the system may change over time, the system is adjustably autonomous. Complex dynamic environments also render advanced planning intractable since the planning process may not terminate before the state of the environment changes. Plan revision, deliberation based on the contents of the plan library, and dynamic rescheduling all tolerate only mild time constraints. This

58

Norms and Adjustable Autonomy

realization has caused many researchers in artificial intelligence to change their view of the environment from modeling and searching the entire state space, to accepting uncertainty with respect to future events. Behavior-based robotics [24] and hybrid architectures [4] are prime examples of this development. Since many of the changes taking place in the environment may be caused by the actions of other agents (human or artificial), the field of multi-agent systems is relevant [22]. Agent action in complex environments is suitably controlled by an anytime algorithm [43]. Alternatively, the agent may turn to a human supervisor. This presupposes the supervisor’s availability and ability to solve the decision situation correctly, and within the time available to her. Internal decision modules can handle simple decision situations. In more complex situations, these modules can be externalized, downsizing the code of the agents. The size of an agent is often pivotal, for example if the agent is mobile, or if it resides in hardware equipped with only a small memory [18]. Such external decision support to artificial agents can be placed in a pronouncer [19], [119]. One can envision an anytime algorithm allowing for the agent to pose the same question to both a human supervisor and a pronouncer. In any case, the availability of pronouncers decreases the need for human supervision. The concept of autonomy is also studied in research on agents. Definitions of levels of autonomy in agent theory differ remarkably from those in the adjustable autonomy literature. We will discuss this in Section 2. The role and of norms in agent decision making will be discussed in Section 3. Conclusions will be presented in Section 4.

8.2

Adjustable Autonomy

Adjustable autonomy is a recently introduced term from the field of autonomous systems in space exploration [47]. The general idea is that autonomous systems adapt the level of autonomy to the situation at hand. Individual agents participate in several groups and may

Norms Can Replace Plans

59

revise their standing toward the group and its members during execution. This enhances their self-organizing capabilities and increases their flexibility and (re)action repertoire. Dorais et al define the level of autonomy as being dependent on: • the complexity of the commands • the number of autonomously controlled subsystems and couplings between these systems • the circumstances that will allow for override of manual control • the circumstances under which user information or control is requested (1999 IJCAI Workshop on Adjustable Autonomous Systems Call for Papers) • the duration of autonomous operation (restated in 1999 IJCAI Workshop on Adjustable Autonomous Systems Call for Papers as the resources consumed, including time) In the field of multi-agent systems (MAS), several authors, e.g., [39], [33], [128], and [44] have developed more or less overlapping hierarchies of autonomy based on the types of processes for decision making that the agents use. The autonomy models developed in MAS research are summarized and extended in [124]. In short, decision making in MAS is made at four separate connected levels: • the level of actions • the level of plans • the level of goals • the level of norms These two definitions of levels of autonomy are quite different. The basic difference lies in their view on autonomy:

60

Norms and Adjustable Autonomy • levels of autonomy as abstraction levels, or the control the agent has over its behavior and decision making process • level of autonomy as level of independence of group (or human supervision)

The work on adjustable autonomy is concerned with the second type of autonomy whereas MAS research is primarily concerned with the first. Reconciling these two is a non-trivial problem. The complexity of the commands (from humans to agents) can be measured in terms of the transfer of control (or delegation) that the communication yields in terms of the levels of actions, plans, goals, and norms. The following example is taken from [47]. Consider the commands: ”find evidence of stratification in a rock” and ”go straight 10 meters”. The first command transfers a goal to the agent, with the agent assuming responsibility to judge what is to count as an adequate result, which object to examine, etc. The second command also transfers a goal but also what is to count as an adequate result. It can even be debated whether this second command is at the level of actions or not. Autonomy in the social space increases with the number of subsystems under agent control since the agent is less dependent on other agents to perform tasks. This is related to the action repertoire of the agent. The need for, and implementation of, adjustable autonomy for governing overriding through human commands is well documented in the agent literature. One domain in which a MAS with relatively autonomous agents must expect frequent human intervention is intelligent buildings [18]. Agent intervention in human activity is less studied but occurs e.g. in surveillance, alarm, and tutoring systems [95]. Decision making at the level of norms (or in more simple cases applying norms) is to decide on overriding manual control or for that matter any outside control [12]. The interaction between the agent and human supervisors (e.g., requests for information) is not modeled as a separate level of autonomy in the agent theory literature. Interaction consisting of information transfer from

Norms Can Replace Plans

61

the human supervisor to the agent is conceptually equivalent to results of information gathering from the environment. Requests for transfer of control is equivalent to the transfer of control stipulated above. The introduction of time as a separate autonomy level (the last point in the list of autonomy levels in the adjustable autonomy literature) is an important addition, although it in fact is merely making explicit the role time plays in the MAS view on autonomy. The levels of decision making are closely linked to time. Both the time-span of execution and the time-span used for self-reflection increase with the levels mentioned. Having control over decision making and behavior over increasing periods of time makes the agents less dependent and thus more autonomous.

8.3

Norms

Norms can, like autonomy, be viewed in different ways. In [12] norms are seen as prescriptive constraints on behavior. In the current work we follow the view presented in [65]. Habermas identifies the use of norms in human action patterns as normatively regulated action. The central concept of complying with a norm means fulfilling a generalized expectation of behavior. The latter does not have the cognitive sense of expecting a predicted event, but the normative sense that members are entitled to expect a certain behavior. This normative model of action lies behind the role theory that is widespread in sociology ([65] p.85, original emphasis). This view is also in agreement with [113]. Tuomela distinguishes between social rules and social norms. Social rules are produced by an authority system. Social norms (or proper social norms in Tuomela’s words) are either conventions or group-specific norms based on mutually shared beliefs about what group members are normatively expected to do. Organizational theory literature, cf. [88], links the

62


use of social rules and social norms as coordination mechanisms to characteristics of the environment. The use of social rules via power structures and authority relations is one way of ensuring group members behave the way they are supposed to. This requires direct supervision and is therefore ill suited for dynamic environments with low interaction rates. The use of social norms is complementary to the use of social rules since it is an intrinsic motivational force instead of an extrinsic one. The use of social norms is helpful when dealing with environments where the interaction rate is low, i.e., where an agent has to be able to trust another agent to a high extent since the possibilities of direct control are limited.

8.4

Conclusions and Future Research

Small and simple distributed active objects with good protocols are what multiagent systems research offers programmers of complex systems. The disadvantage is that the cooperation of the distributed agents is complex and unpredictable. Another solution is to put humans in the loop. This is the solution adjustable autonomy research opts for. This presupposes the availability of humans and requires adequate decision-making capabilities. We propose that the complexity of the system be instead reduced by the removal of sophisticated planning capabilities, since these rely on unrealistic assumptions about the nature of the environment. A hybrid architecture that includes pronouncers may replace human supervision in autonomous systems, at least to the extent that is feasible for political reasons. Pronouncers can enhance agent behavior in systems with onboard reactive planning, but can also be used in systems without any planning. In such systems, norms can be used to coordinate the behavior of autonomous agents.

Chapter 9 Adjustable Autonomy, Norms and Pronouncers Abstract Discussions of agent autonomy levels largely ignore the function of norms in the agents’ decision making process. Allowing artificial agents to use norms makes smooth interaction in the social space less problematic. Agent autonomy will also increase if one grants agents normative support, making accurate and consistent models of other agents feasible. This will result in better behavior predictions and simulations of not only physical but also social systems.

9.1

Introduction

The description of degrees of artificial agent autonomy is problematic from a conceptual point of view. In the field of multi-agent systems (MAS), several authors, e.g., [39], [33], [128] and [44] have developed more or less overlapping hierarchies of autonomy based on the types of processes for decision making that the agents use. This problem will be dealt with in Section 2 below. Our focus will be on control, distributed between human and artificial agents belonging to the same MAS. We will refer to a MAS populated by humans as well as artificial agents, in which the agents model all other agents that they are aware of, as a social space. One aspect of agent autonomy is largely ignored in the current

64


literature, viz. the level of norms. Designing agents that respect social norms is not only helpful in achieving rational agent behavior, but also enhances the possibility of using accurate models. We will discuss the function and learning of norms and also the results of a simulation study of the learning of norms in Section 3. Norms can also be used by external counseling devices such as pronouncers [17]. A pronouncer is an entity providing normative advice to agents. It is external and usually not wrapped into an agent itself. By contrast, internal procedures are usually called decision modules. The use of pronouncers by idle or disoriented agents reduces the need for human intervention. It also makes the social space less dependent on any outside control, such as an earth-based mission control center [47]. Pronouncers are discussed in Section 4. The need for, and implementation of, adjustable autonomy for governing overriding through human commands is well-documented in the agent literature. One domain in which a MAS with relatively autonomous agents must expect frequent human intervention is intelligent buildings [17]. Agent intervention in human activity is less studied but occurs e.g. in surveillance, alarm, and tutoring systems [95]. Requirements on the simplicity of the agents involved in such systems have so far kept designers from implementing sets of norms. We indicate below why this is likely to change in future designs, and present our conclusions in Section 5.

9.2

Simulation of Learning of Norms

In human societies, norms have a dual role in that they both serve as filters for unwanted actions, plans, or goals and as help in predicting the behavior of other members of society. Implemented norms for artificial agents can be used for predicting the behavior of humans in the social space, and thus helps agents maintain a good domain model. It will also enable the agents to become reactive: they can recognize other agents as agents instead of just objects. The use of norms in

Adjustable Autonomy, Norms and Pronouncers

65

artificial agents is a fairly recent development in MAS research (cf. e.g., [102], [124], [12]). The learning of norms can be divided in two types, viz. the emergence of norms [115] and the acceptance of norms [40]. These two types of learning express learning at different levels. The emergence of norms is learning at the level of the social system while the acceptance of norms is learning at the level of the individual agent. In [40] reasons for the acceptance of norms are discussed. We are not primarily interested in why agents accept norms since we presuppose that membership of a coalition implies the agents accept the norms of the coalition. Instead we are interested in how acceptance of norms changes the decision making behavior of the agents by changing the agents definition of the norms of the coalition (norm-spreading) and by the adaption of the agents’ own norms (norm-internalization). We have conducted simulation studies of the spreading and internalizing of norms as a function of autonomy towards the coalition (at the level of actions). Agents forming a coalition roam around in a two-dimensional world where two resources are randomly distributed. Agents have two decision trees, one containing the subjective evaluation of all alternatives (i.e., a self-model) and one containing the subjective view on the coalition’s evaluation of all alternatives (i.e., a coalition model). An agent chooses one of the alternatives and tells the other agents about its choice. Feedback of these agents is used to update the coalition model. The self-model is updated based on the feedback from the environment (i.e., if the chosen alternative is realized or not). Deciding which alternative to choose entails balancing the self-model and coalition model. If the agent is completely autonomous with respect to the coalition, it only evaluates its selfmodel. If an agent has no autonomy with respect to the coalition, it uses its coalition model. To be able to use the coalition model to predict the behavior of other agents, each agent should have the same coalition model (i.e., share the norms of the coalition). We thus measured the spreading of norms in the coalition.

66


The spreading of norms is measured as the differences in the coalition utility bases over the agents (i.e., the mean value of the standard deviation per alternative of the coalition utility of that alternative of each agent). We can imagine two situations in which agents totally comply with the group norms. The agents may have no autonomy, or the agents may have adapted to the coalition model to the extent that their self-model equals the coalition model. For this purpose we measured the internalizing of norms as the difference between an agent’s own utility base and the coalition utility base it has. A hypothesis was formulated: Hypothesis: the higher the degree of autonomy, the lower the predictability of behavior will be. The simulations showed that an increase of the agents’ autonomy resulted in a decreased norm-spreading. The norm-internalizing factor did not have such a straightforward relationship as we had hypothesized. We suspect that this is due to the second-order type of learning involved in the internalizing of norms. Further simulation studies [121] will be conducted to clarify this.

9.3

Pronouncers

When an intelligent agent has to decide on what action to take, it might ask for advice. The base case is the agent asking itself what to do next. The even more difficult case is when the precarious agent asks someone (or something) else. This case can in turn be analyzed by considering two sub-cases. Firstly, the agent may ask other agents in its MAS. This situation can often be reduced to the base case, viz. if one assumes a fully functioning communication architecture for co-operating agents. Second, the agent may consult an entity outside 4the MAS that might not be an agent at all. This entity may come in different guises, e.g., a human, a blackboard, or an oracle. Such entities have too many variations to allow for them


67

to be studied in precise terms: a blackboard, for instance, does not entail the same agent architecture or model to all researchers that claim to use them. The entity might at times be inaccessible to the querying agent, and the entity data indeed accessible to the querying agent is usually incomprehensible to the agent. The standard way to overcome this is to use a wrapper [57], but the size and complexity of the wrapper code for an entity of the kind we study is unacceptable in domains with noticeable time constraints [131]. The agent in need of advice may instead feed a pronouncer with a description of a decision situation, including its subjective assessments of relevant utilities and probabilities. How the pronouncer accesses these assessments is not important for our discussion. The pronouncer also has access to a norm base containing all norms. Each agent coalition has its set of norms, a subset of the norm base, and an agent can belong to many coalitions. Regardless of the coalition structure, the agents turn to the same pronouncer for advice. It is the involvement of norms that makes a pronouncer more than just a decision rule aimed at maximizing an agent’s expected utility. The norm base can be used to disqualify agent actions, plans, or goals if they fail to adhere to the norm set applying to the agent. It can also be used for calculating punishments and rewards, if agent feedback is used. Naturally, one can imagine a simple MAS in which each agent has the same responsibility towards a group. Then it would suffice to store norms globally, as part of the pronouncer. The realistic and most general case, however, is where each agent has unique obligations towards each and every one of the other agents. For instance, a MAS might consist of 200 agents in which a particular agent has obligations towards the entire population (including itself), but also towards two overlapping strict subsets of, say, 20 and 25 agents that constitute coalitions. These coalitions might be dynamically construed, something which will affect the nature of obligations heavily over time. The control of coalition membership for human agents is different

68


from that for artificial agents, as are the reasons for choosing to join a coalition. In the intelligent building MAS described in [16], for instance, a so-called personal comfort agent might be instructed by its human owner to join a coalition of agents representing people working on the same floor, or to join a coalition of agents that have the same temperature and lighting preferences for the conference room. A particular personal comfort agent may then join other coalitions on different grounds, e.g., to get a better bargaining position in negotiations regarding the lighting in the conference room. The human agent owning the agent will not be informed of this rational step taken by the agent. In fact, the human agent typically feels strongly that he or she should be kept out of such negotiations, and is willing to grant his or her personal comfort agent enough autonomy for it to remain almost invisible. Since socially intelligent behavior is the goal of one part of the adjustment of autonomy in [47], viz. a request for advice from an artificial agent to a human agent, the use of pronouncers may reduce the need of human interference. The same goes for possible requests for advice from any type of agent to any other type of agent. Pronouncers may also be used to replace the need for input from human agents in interactive planning. In both cases, the artificial agents will be more autonomous with respect to the human agents. The use of pronouncers in a simulated robotic soccer team has been implemented [131] and will be further developed for the 1999 RoboCup world championships, in both a simulated and a physical legged-robot team. Robotic soccer is a real-time environment with incomplete information at the level of the agent, making it an ideal testbed. An important indicator concerning adjustable autonomy is time constraints. The amount of time available is pivotal to whether or not a pronouncer call can be made. In [131], it is demonstrated that pronouncer calls can be beneficiary even under the dynamic real-time constraints of RoboCup.


9.4

69

Conclusions

Allowing artificial agents to use norms as constraints on actions, and also for enriching their domain models with respect to the groups they act within, is necessary for smooth interaction with humans and with other artificial agents belonging to the same social space. Equipping agents with normative decision making features, together with the ability to make pronouncer calls, makes the agents more autonomous. The agents may then have more accurate and consistent models of each other, thus enabling better behavior predictions and simulations of not only physical but also social systems.

70


Chapter 10 Adjustable Autonomy, Delegation and Distribution of Decision Making Abstract In this paper we will illustrate the possible uses of norms in multi-agent decision making. Norms as obligations serve to generate action repertoires and norms as constraints on behavior enable pronouncers to filter out subrational decisions. These are combined with adjustable autonomy to produce highly adaptive agents. The theoretical framework is tested in simulations of norm-spreading and norm-internalization and also robotic soccer.

10.1

Introduction

The focus of this paper is on how agents may obtain a balance between following the norms of the coalition and individualistic decision making (i.e., maximizing personal profit). Since agents may act in dynamic environments, the choice between these two extremes should not be fixed beforehand. In other words, agents should be able to adjust their autonomy with respect to the coalition at run time. Another interesting issue is how the set of shared norms comes about. Adjustable autonomy is a recently introduced term from the field of autonomous systems in space exploration [47]. The general idea is

72


that autonomous systems adapt the level of autonomy to the situation at hand. One aspect of agent autonomy is largely ignored in [47], viz. the level of norms. Designing norm-regulated or normautonomous agents will enhance the possibilities of accurate models of humans for the agents and vice versa. The need for override of human commands by the artificial agents can be modeled more clearly when using the concept of norms. Furthermore, norms can be used by external counseling devices such as pronouncers [13]. Agents finding themselves in a situation too complex for them to make a rational decision may ask an external entity for advice. A special case of such an entity is a pronouncer. A pronouncer is an artificial decision maker that is not an agent nor part of the coalition. Pronouncers take as input the decision tree of the agent, a description of the situation and the set of norms of the agent and has as an output the most rational choice given the information provided. The use of pronouncers for advice will reduce the dependence of agents on other agents (human or artificial). It will also enable the artificial and human agents together forming a society to be less dependent on any outside control. Individual agents participate in several coalitions and may revise their standing toward the coalition and its members during execution. This enhances their self-organizing capabilities thus increasing their flexibility and (re)action repertoire. First we will look more closely at the concept of autonomy before discussion ways of delegation and distribution of decision making. After this we will introduce the two simulation models developed to test the usability of these concepts in multi-agent systems and discuss the results obtained so far. Finally we will indicate possible topics for further research.

10.2

Simulation of Norm-Spreading and NormInternalization

The simulation model consists of several agents roaming a two dimensional space. The agents form a coalition with one of the agents acting

Delegation and Distribution of Decision Making

73

as the leader. Every spot in the two dimensional space may contain either nothing, one piece of resource one, one piece of resource two or one piece of both resources. The agent has a choice to either do nothing, move to another spot or take resource one or resource two (if available). Combining the number of content alternatives with the choice alternatives and outcome alternatives (if the chosen alternative is realized or not) gives 20 combination alternatives in total. 10.2.1

Description of Decision Making Model

Every agent has a private utility base containing utility values for each of these alternatives (self-model) and a coalition utility base containing the utilities the agent presumes the coalition has for each of the alternatives (coalition model). The coalition model expresses the agent’s interpretation of the norms the coalition holds. The degree of autonomy of an agent relative to the coalition determines to what extent the coalition model is followed when making a decision. E.g., an autonomy of 0.4 expresses that in 40 percent of the cases the coalition model is followed instead of the self-model to make a decision. Choosing an alternative does not mean this is also bound to happen, chance and the agent’s skills influence the outcome. An agent updates its self-model and coalition model based on the outcome and the result of the feedback coalition members give in answer to the agent’s message containing the chosen alternative. The influence of the leader of the coalition may also vary. The leadership value expresses the weight of the information provided by the coalition leader. E.g., a leadership factor of 2 expresses that the feedback from the leader is twice as important as the feedback of the other coalition members. 10.2.2

Simulation Setups

The following simulations were run: autonomy (on a scale from 0 to 1) had a value of 0.0, 0.4 or 0.8 and the leadership value was 1, 2, 5 or 10. In total this gives 12 simulation setups. The following two hypotheses were formulated:

74

Norms and Adjustable Autonomy • Hypothesis 1: the higher the degree of autonomy, the higher the variance of behavior will be. • Hypothesis 2: the higher the leadership value, the lower the variance of behavior will be.

The variance of behavior can be measured in several ways. One way is by determining the difference between an agent’s own utility base and the coalition utility base it has. This expresses the norminternalization. Another measure is the differences in the coalition utility bases over the agents. This expresses the norm-spreading. The spreading of norms is graphically displayed as the mean value of the standard deviation per alternative of the coalition utility of that alternative of each agent. 10.2.3

Implementation of the Simulation Model

The simulation model is implemented in Java. Each agent is a separate thread. The agents communicate with the environment and each other through a router programmed using JATLite. Varying the settings for the agents requires editing of some datafiles and Java code files and compiling these. All simulation runs were run for 100 minutes. During the simulation run a log file is kept for each agent which gets updated every minute with the agents self-model and coalition model at that point in time. 10.2.4

Simulation Results

The results of some of the simulations are shown in the figures 1 to 4. Figure 1 shows the behavior of the norm-spreading factor in time with an autonomy of .4 for the three different leadership values. Figure 2 shows the influence of varying the autonomy factor while the leadership value is held constant at 2. Figure 3 shows how the norminternalization factor behaves in time for the various leadership factor settings while the autonomy factor is set to .4. Finally, figure 4 depicts


75

the norm-internalization factor over time for a leadershipfactor of 2 and varying the autonomy factors.

Figure 10.1: The norm-sharing factor with autonomy = .4.

Figure 10.2: The norm-sharing factor with leadership = 2.

10.2.5

Interpretation of the Preliminary Results

Both figure 1 and figure 2 comply with the formulated hypothesis. One may observe that increasing the autonomy causes the agents to need more time to reach the maximum value of the norm-sharing value and it also takes more time for the agents to enter a stable situation.

76


Figure 10.3: The norm-internalization factor with autonomy = .4.

Figure 10.4: The norm-internalization factor with leadership = 2. Increasing the leadership factor on the other hand makes the agents reach the norm-sharing value’s maximum sooner and also get to a stable situation sooner. The norm-internalization factor graphs as depicted in figure 3 and figure 4 do not comply with the formulated hypothesis. Several factors may play a role here. One possible cause is that not all situations occur during the simulation. Since the norm bases are only updated for the situations that occur, some utilities do not change during the entire simulation. A second explanation may be that the variance between different runs with the same setting could


77

be greater then the difference between runs with different settings.

10.3

Related Work

An example of research on obtaining a balance between deliberations at macro and micro levels is [70] in which a proposal for social rationality is developed. The authors try to give a general decision rule based on payoff functions (in terms of benefits and losses) for individuals and society. This group decision rule is however not resistent to individual differences between the agents (i.e., all agents are supposed to follow the group rule) and agents can be member of only one group (i.e., the whole agent systems forms one social system which has no subsystems). As mentioned previously, the macro-micro link problem and the function of norms is discussed in [39]. In more recent work [40] a model for norm acceptance is developed. In a sense this work is complementary to our work. We do not study the acceptance of norms but presuppose that coalition membership in itself is reason to accept norms. On the other hand, unlike Conte et al. we do not presuppose that accepting norms implies following them. In the work on Sugarscape [50] some form of norm spreading is also studied. The model of cultural transmission is too basic and its transmission rule too general to be considered a model of norm-spreading reconcilable with deliberating agents. Peter Stone, who recently received his Ph.D., at the Carnegie Mellon University (CMU), has in co-operation with professor Manuela Veloso and Patrick Riley developed a team, CMUnited, which became the champion of the World Cup 1998. Stone’s research is more focused on machine learning [110] than on the social aspects of RoboCup. Milind Tambe and Gal Kaminka of the University of Southern California, have used a rule-based framework approach when developing their team ISIS [111] in Soar. Tambe and Kaminka do research in the field of teamwork and social diagnosis, which is not far from the research on social norms performed in our lab [19] and described in this paper.

78


10.4

Discussion and Future Research

10.4.1

Simulation of norm-spreading and norm-internalization

Further simulation experiments will be conducted to draw conclusions about why the norm-internalization factor does not comply with the formulated hypothesis. Other future work in the simulation of normspreading will include comparison of different initial situations (i.e., a default coalition norm base versus an initial coalition norm base being equal to the self norm base). Another topic for future research will be the formation of coalitions. Previous work on this topic has been focused on game theory and thus individualistic decision making. Inspiration from social theory (e.g., [59], [61], [113]) will enrich these models with normative decision making. 10.4.2

RoboCup

We will continue to develop our simulator league team in the forthcoming years and we will implement the main concepts mentioned in this paper. We will also investigate whether it is beneficial to use several groups and leaders within a team and to which extent the agents should be autonomous related to the group.

Chapter 11 On the Learning of Norms Abstract Reducing behavior variance among the agents can be achieved by either implementing an organzational structure or via the sharing of norms. Norms as a coordination mechanism preserves the agents autonomy to some extent. Establishing a shared norm set under run time can be problematic. This paper examines one mechanism, viz. the communication of norms after decision making such that agents do not have to await feedback. One issue is of major concern, which is the norm set agents use initially. The effect of varying agent autonomy and initial norm set on the spreading and internalizing of norms was examined in a simulation study. It was found that the spreading of norms is independent on the bootstrap procedure choosen whereas the internalizing of norms is highly dependent upon it.

11.1

Introduction

Bridging the gap between micro and macro level behavior is one of the main goals of the work of the research group of Castelfranchi et al. (see e.g., [39], [40]). Of central concern in the work of this group is the cognitive anchorage and learning of norms. I will here focus on the second part, viz. the adaptation of individual norms and the evolution of group norms from individual norms. Thus, I study the relation between micro and macro levels of constraints on behavior

80


in both directions. I will not concern myself with issues of group formation and the mechanisms underlying the entering and leaving of groupmembers since this will be the focus of future work. Norms have a double functionality, they steer the agent’s behavior and serve as models to predict the behavior of fellow group members. In this paper I will view norms as basically being constraints at the level of actions, and I will ignore the role of norms in the creation and selection of goals or plans. This makes the agents norm-regulated rather than norm-autonomous [123], [40]. The modelling of norms and the learning of norms is partially based on [12]. In [12], a general model for artificial decision making constrained by norms was presented. In this model agents adhere to norms via local adaptation of behavior or via groups exercising their right to disqualify action options. The adaptation of behavior consists of an internalization of group norms, or more precisely a synchronization of the individual norms to those of the group. The learning of norms constitutes to an individual behavior pattern endorsed by the group and is thus the basis for socially intelligent behavior. The assessments in the information frames gradually evolve, in order for the agent to act in accordance with the norms of its group. The group norms, or social constraints, are not merely the union of the local information frames of its members, but rather develop interactively, as do the local information frames. In following the functional theories of Parsons [92], organizational theory literature (cf. [88]) distinguishes two mechanisms for the coordination of agents. Organizational structure1 or more general the use of power structures and authority relations is one way of ensuring group members behave the way they are supposed to. This requires direct supervision and is therefor not suited for dynamic environments with low interaction rates. The second type is the use of organiza1

In [88], organizational structure is defined as “the use of centralization of decision-making, formalization of rules, standardization of work processes and skills, and/or control of output by acceptance of only adequate outcomes” (p.3).

On the Learning of Norms

81

tional culture2. Thus, organizational culture is complementary to the use of organizational structure since it is an intrinsic motivational force instead of an extrinsic one. The use of shared culture (i.e., norms) is helpful when dealing with environments where the interaction rate is low, i.e., where an agent has to be able to trust another agent to a high extent since the possibilities of direct control are limited. The use of norms as a mechanism for behavior variation reduction assumes at least two components: a theory of acceptance of norms (which is the focus of [40]) and a mechanism for the spreading and internalizing of norms. I will focus on the second component and in particular test the possible use of communication of normative advice as a way of spreading norms. As for the acceptance of norms, I presume that if an agent is part of a group it blindly accepts the norms of that group (one might consider this as a benevolence assumption [97] at the normative level). The degree to which the group norms are applied in the agent’s decision making is dependent upon the degree of autonomy the agent has with respect to the group. More precisely I will study the influence of normative comments on previously made choices. This contrast with the normspreading mechanism in [12] where normative advice is sought before an agent makes its choice. First I will look more closely at the concept of autonomy before discussion ways of delegation and distribution of decision making. After this I will introduce the simulation model developed to test the usability of these concepts in multi-agent systems and discuss the results obtained so far. Finally I will indicate possible topics for further research. 2

In [88], organizational culture is defined as “consensual schema shared among employees in an organization, resulting from a pattern of basic assumptions and norms enhancing individual and organizational stability, manifested in shared meanings, communicated by stories, myths, and practices, and resulting in certain behavior patterns which are unique to the organization” (p.5).

82

11.2


About Norms

In human societies, norms have a dual role in that they both serve as filters for unwanted actions, plans, or goals and as help in predicting the behavior of other members of society. Implemented norms for artificial agents can be used for predicting the behavior of other agents in the social space (be they human or artificial), and thus helps agents maintain a good domain model. It will also enable the agents to become reactive: they can recognize other agents as agents instead of just objects. The use of norms in artificial agents is a fairly recent development in MAS research (cf. e.g., [102], [124], [12]). The learning of norms can be divided in two types, viz. the emergence of norms [115] and the acceptance of norms [40]. These two types of learning express learning at different levels. The emergence of norms is learning at the level of the social system while the acceptance of norms is learning at the level of the individual agent. In [40] reasons for the acceptance of norms are discussed. I am not primarily interested in why agents accept norms since I presuppose that membership of a group implies the agents accept the norms of the coalition. Instead I am interested in how acceptance of norms changes the decision making behavior of the agents by changing the agents definition of the norms of the group (norm-spreading) and by the adaption of the agents’ own norms (norm-internalizing).

11.3

Simulation of Spreading and Internalizing of Norms

The simulation model consists of several agents roaming a twodimensional space. The agents form a group with one of the agents acting as the leader. Every spot in the two dimensional space may contain either nothing, one piece of resource A, one piece of resource B or one piece of both resources. The agent has a choice to either do nothing, move to another spot or take resource one or resource two


83

(if available). Combining the number of content alternatives with the choice alternatives and outcome alternatives (whether the chosen alternative is realized or not) gives 20 combination alternatives in total. 11.3.1

Description of Decision Making Model

Every agent has a private utility base containing utility values for each of these alternatives (self-model) and a group utility base containing the utilities the agent presumes the group has for each of the alternatives (group-model). The group-model expresses the agent’s interpretation of the norms the group holds. The degree of autonomy of an agent relative to the group determines to what extent the group-model is followed when making a decision. E.g., an autonomy of 0.4 expresses that in 40 percent of the cases the group model is followed instead of the self-model to make a decision. Choosing an alternative does not mean this is also bound to happen, chance and the agent’s skills influence the outcome. An agent updates its self-model and group model based on the outcome and the result of the feedback group members give in answer to the agent’s message containing the chosen alternative. 11.3.2

Simulation Setups

The following simulations were run: autonomy (on a scale from 0 to 1) had a value of 0.0, 0.4 or 0.8 and the initial group-model was either set to a default model (equal for all agents) or equal to their self-model (different for all agents). This gives 6 simulation setups in total. The following two hypothesis were formulated: • Hypothesis 1: the higher the degree of autonomy, the higher the variance of behavior will be. • Hypothesis 2: if the self-model equals the group-model, the agents will adapt their group-model at a slower rate than when

84

Norms and Adjustable Autonomy the self-model and group-model differ and thus will have a higher variance of behavior.

The variance of behavior can be measured in several ways. One way is by determining the difference between an agent’s own utility base and the group utility base it has. This expresses the internalizing of norms. Another measure is the differences in the group utility bases over the agents. This expresses the spreading of norms. The spreading of norms is graphically displayed as the mean value of the standard deviation per alternative of the group utility of that alternative of each agent.

11.4

Implementation of the Simulation Model

The simulation model is implemented in Java. Each agent is a separate thread. The agents communicate with the environment and each other through a router programmed using JATLite. Varying the settings for the agents requires editing of some datafiles and Java code files and compiling these. All simulation runs were run for 100 minutes. During the simulation run a log file is kept for each agent which gets updated every minute with the agents self-model and groupmodel at that point in time.

11.5

Simulation Results

Both figures comply with the formulated hypothesis. One may observe that enlarging the autonomy causes the agents to need more time to reach the maximum value of the normsharing value and it also takes more time for the agents to ante a stable situation. The norm internalization factor graphs as depicted in figure one and figure 2 do not comply with the formulated hypothesis. Several factors may play a role here. It could be that the system needs more time to enter a stable situation then the amount of time the simulations were


85

Figure 11.1: The normspreading factor for all simulation setups

Figure 11.2: The norminternalizing factor for all simulation setups

86


run. Another possible cause is that not all situations occur during the simulation. Since the norm bases are only updated for the situations that occur, some utilities do not change during the entire simulation. A third explanation may be that the variance between different runs with the same setting could be greater then the difference between runs with different settings. Further simulation experiments will be conducted to draw conclusions. Both figure 1 and figure 2 comply with the formulated hypothesis. One may observe that increasing the autonomy causes the agents to need more time to reach the maximum value of the norm-sharing value and it also takes more time for the agents to enter a stable situation. The norm-internalization factor graphs as depicted in figure 3 and figure 4 do not comply with the formulated hypothesis.

11.6

Related Research

First steps in the direction of defining rational behavior as individually rational behavior extended by collective awareness, i.e. the micromacro link, were taken within the MAS area by Castelfranchi and his group [39]. The beginnings of a formalization of norms in agent action was through a logic approach [126]. The most popular logic is deontic logic, usually appearing as revised versions of Meyer’s reduction of deontic logic to dynamic logic [82]. Only actions, and not formulas (representing assertions or assessments) can be obliged in Meyer’s logic, and this has led to various extensions. These include (roughly in order of sophistication) a coupling to speech acts using illocutionary logic [45], a “logic for action and norms” [46], and even first steps towards a logic-based social agent development language [6]. Interestingly enough, the hard work put down to augment Meyer’s logic is debatable in view of our somewhat controversial view of plans as normative advice for essentially reactive agents. Our view is in line with Meyer’s original position that world states need not be explicitly


87

modelled: Modelling actions is sufficient. Jennings and Campos refine Newell’s principle of rationality [85] into a “principle of social rationality” that says that “if a member of a responsible society can perform an action whose joint benefit is greater than its joint loss, then it may select that action” [70]. Since it is modelled on Newell’s principle, it inherits some of its weaknesses, e.g., it does not treat the case of several (or no) suitable actions. Moreover, the central concept of joint benefit is defined for the coarsest possible value scale, viz. one of loss and benefit only. In a paper by Kalenka and Jennings [71], benefits are divided into individual, social, and joint benefits. As in [70], evaluations are based on utility functions representing agent preferences. These two papers are important first steps towards a social level. The natural second step is to introduce beliefs (i.e, probabilities), vagueness (i.e. imprecise utility assessments), and a more expressive language, including risk profiles, security levels, individual constraints towards groups, etc.

11.7

Discussion and Future Research

Several factors may play a role here. One possible cause is that not all situations occur during the simulation. Since the norm bases are only updated for the situations that occur, some utilities do not change during the entire simulation. A second explanation may be that the variance between different runs with the same setting could be greater then the difference between runs with different settings. Further simulation experiments will be conducted to draw conclusions about why the norm-internalizing factor does not comply with the formulated hypotheses. Another topic for future research will be the formation of groups. Previous work on this topic has been focused on game theory and thus individualistic decision making. Inspiration from social theory (e.g., [59], [61], [113]) will enrich these models with normative decision making.

88


Part III Simulations of Organizational Problem Solving

Chapter 12 TASCCS: A Synthesis of Double-AISS and Plural-Soar Abstract Double-AISS [81] was possibly the first completed effort to build an artificial intelligence (AI) based model of organizational decision making. Intended as a follow-up study of the garbage can model of organizational decision making [37], Double-AISS examined the influence of various organizational parameters on the problemsolving behavior in organizations. In 1990, the Soar architecture [74] was adapted to build another AI-based model of organizations, called Plural-Soar [32]. Plural-Soar examined the influence of communication and learning capabilities on the problem solving behavior of independent actors working on a simple task. The model reported in this chapter, TASCCS provides—in a sense—a synthesis of Double-AISS and Plural-Soar.

12.1

Double-AISS

Double-AISS [81] provides a generalization of the garbage can theory of decision making [37]. “Garbage can” is a metaphor used to characterize loosely coupled decision processes: Preferences are unstable, the technology is unclear, and participation is fluid. In the garbage can model, a decision is an outcome of several relatively independent processes within an organization. Double-AISS tries to overcome the

92


drawbacks of earlier renderings of the garbage can model [2], [28], [91] by using AI techniques rather than numerical equations. The core of the model is formed by the actors. They make the decisions in the model. Actors are embedded in an organizational structure, which maps the communication possibilities onto the actors. The content of the communication is an issue. What actors can do with issues is defined by their skills and actions. The acronym Double-AISS is derived from these five building blocks. 12.1.1

The Model of an Actor in Double-AISS

In Double-AISS, the rationality of actors is bounded [104]. They: • Do not know all decision alternatives in advance. • Do not anticipate all consequences of their actions. • Do no try to optimize but rather to satisfice (that is, try to meet some aspiration level, which depends on prior experience). • Do no have completely ordered preferences (whereas their commitment to each other or to the organization may be limited). The organizational structure is implemented as a communication network, which defines which actors can communicate with each other. The content of the communication is an issue. Issues are multidimensional sets of interrelated subtasks. In the Double-AISS, the categorical task is the production of a memo. This task consists of six subtasks: writing, drafting, typing, editing, approving and copying. Every subtask is a dimension of the memo task. Skills are the qualifications of an actor. Skills are “applied” to problems, so if an actor can draft a memo and has an issue that contains the need for drafting, then the actor can draft the issue, with the drafting task completed, the issue’s subtasks are reduced to five. An actor can choose between several strategies to solve the memo task. The possible actions are reducing an issue, moving an issue

A Synthesis of Double-AISS and Plural-Soar

93

to another actor, attracting an issue from another actor, combining two issues into one new issue (which has less dimensions than the two old issues had together), splitting one issue into two new issues (which have more dimensions together than the old one had), or doing nothing. 12.1.2

The Search Space of an Actor in Double-AISS

Double-AISS’s organization consists of 10 actors. They make their decision as a function of search through their own problem space. The path of the search depends on seven factors: (a) the search strategy, (b) cognitive capacity, (c) aspirations, (d) preferences, (e) organizational work load, (f) commitment, and (g) organizational structure. 1. The search strategy can be either of the following two: (timerelated) depth-first search of (time-related) breadth-first search. In depth-first search, an actor tries to anticipate the future by conceiving sequences of activities. In the process. the actor may try to imagine what another actor would do, given that the alternative under consideration is chosen. In breadth-first search, an actor tries all possible alternatives in turn. Both search modes search until either a satisfying alternative is reached, or the cognitive capacity of an actor is exhausted. The choice for a search alternative is made according to a preference for one of the search modes. If the preferred search strategy gives no result, then the other strategy is given a try. If no satisfying alternative is found, then the last-found tolerable alternative is chosen. If there also is no tolerable alternative, then the action “do nothing” is chosen. 2. The cognitive capacity of an actor is the number of search steps an actor can make per decision cycle. 3. The aspiration level of an actor is the number of tasks an actor is the number of tasks an actor expects to complete (“reduce”)

94

Simulations of Organizational Problem Solving in one decision cycle. An actor also has a tolerance level. If the aspiration level is not met when the time has come to make a decision, but the tolerance level is met. the corresponding action is chosen. 4. The preferences of an actor concern various parts of the decision process. There are preferences for subtasks, skills, actions, other actors, search strategy, and decision outcome (whether or not to reduce one’s own work load at the expense of other actors’ work load). Preferences are implemented via the probability with which an alternative is chosen. 5. The work load of the organizations is the number of issues that come into the organization at one decision cycle, multiplied with the number of tasks per issue. The work load of an actor is its part of the total work load of the organization. 6. The commitment of an actor to the organization can be one of the following two alternatives: The actor is concerned with only its own work load, or it is concerned with the work load of the organization as a whole (also called individualistic versus cooperative actors [8]). The commitment of an actor determines the ordering of the action alternatives. An egotistic actor (concerned only with its own work load) conceives of moving an issue to another actor as a solution, because its own work load decreases. Conversely, attracting an issue from another actor is a non-solution for it. Altruistic actors think of both moving and attracting as a non-solution because the overall work load is not reduced. 7. The structure of the organization determines the communication possibilities of an actor. Communication is effected via the exchange of issues.

A Synthesis of Double-AISS and Plural-Soar 12.1.3

95

The Behavior of Double-AISS

The behavior of the organization in Double-AISS results from the interactive effect of individual decisions. The canonical simulation experiment with Double-AISS [81] examined the impact of various parameters on several measures of organizational performance. The experiment was done as a block design with two variables per block. The independent variables were cognitive capacity, work load, structure, aspiration level, adaptation period, the strength of various preferences, and the maximum depth of the search. The dependent variables were productivity, the number of solutions, non-solutions, nondecisions, and the organizational climate. Cognitive capacity and commitment (egotistic/altruistic) turned out to dominate the behavior of the model. Most other independent variables played a significant role as well, but their influence was of secondary importance. 12.1.4

Critique on Double-AISS

Subsequent work on Double-AISS indicated several points of possible improvement, such as (a) the implementation of subtasks and skills, (b) the reduce calculus, (c) the communication structure and (d) foresight. 1. Both subtasks and skills in Double-AISS are implemented as characters, which serve as abbreviations of the task itself. For example a stands for “approve”, which is both a subtask (as dimension of an issue) and a skill (as a qualification of an actor). This representation may give rise to confusion. Also, subtasks are not related to a subject - for example, one cannot know if a stands for “approve memo X” or “approve memo Y”. Therefore, it is not possible to control the order in which subtasks are carried out, although this order might be important. Approving memo X and typing memo Y do not have to be carried out in a specific order, but approving and typing memo X do.

96

Simulations of Organizational Problem Solving 2. The “reduce” calculus used in Double-AISS is idiosyncratic. In the garbage can theory, the role of an issue (problem or solution) depends on the context. In Double-AISS this context dependency is implemented by means of symbolic valences (minus signs): A minus sign in front of a subtask indicates a solution; a subtask without the minus sign indicates a problem. For example, it is difficult to grasp what it means to have the skill a (which stands for a memo that needs approving). The reduce action effects the combination of a problem and a corresponding solution, with - a as a skill and a as the subtask for an issue. The same context dependency could better be conceptualized via preferences. For example, the preference to “approve”, if strong enough, would put a strain on the decisionmaking process, leaving other tasks “on hold”. This way, skills could represent solutions and issues could represent problems. 3. The communication in Double-AISS is not communication in the ordinary sense, that is flow of information, Instead, communication is effected via the flow of issues. 4. Foresight in Double-AISS is one-sided. The actor who is trying to conceive the future visualizes the possible reaction of other actors by “stepping into their shoes” and not communicating directly with them. And when stepping into another actor’s shoes it has a perfect model of that other actor, which is not realistic.

12.2

Plural-Soar

Soar [74] is an implementation of an elaboration of the human problem-solving paradigm [86]. It is system capable of general intelligent behavior and can perform the full range of problem-solving methods, and learn about all aspects of the tasks and its performance on them. Plural-Soar [32] is a model of a small organization imple-


97

mented in Soar. As in Double-AISS, the model’s core is formed by actors (agents in Soar terminology). The name Plural-Soar is derived from the contraction the contraction of plural agents and Soar. 12.2.1

The Model of an Actor in Plural-Soar

The organization structure is flat—all actors are on the same hierarchical level. Actors have the ability to communicate and to pose questions to all other actors concerning the location of items. The original actor continues searching, whereas the other agents may respond at a later date (if they encounter the item) so one could say a particular problem is solved in parallel mode by the organization, once it is communicated to others. Once the answer is found, it is communicated to the original actor. The task performed in Plural-Soar is a warehouse task. The warehouse receives orders, and the actors have to take an order, find the ordered item in the warehouse, and put it on a conveyer belt. Every actor can work autonomously; that is, it can perform all actions to complete the warehouse task. These actions are: 1. Move to the left 2. Move to the right 3. Move item to left stack 4. Move item to right stack 5. Move item to conveyer belt 6. Ask order 7. Ask question 8. Answer question 9. Wait

98


The choice of any of these actions depends on the state of the actor. For instance, in order to put an item onto the conveyer belt, the actor must be in front of a stack and the top item of the stack must be the ordered item. If there is more than one applicable action, preferences determine the choice of the action alternative. The physical properties of the warehouse or the actions of other actors may cause actors to wait for their turn. Only one actor can access a particular stack at a time. 12.2.2

The Actor’s Search Space in Plural-Soar

The organization consists of (at most) five actors. Implemented as a separate Soar program, every actor is autonomous in its choice of action alternatives. Every actor has a mental model of the warehouse and knows the number of stacks, their location, and the location of the conveyer belt. The search through the problem space is influenced by the behavior of other actors only when they have to wait at a queue or when there is a question pending. When an actor has filled its order, it goes to the order stack to get a new one. Every item is unique, is ordered only once, and is only once available in the warehouse. The warehouse itself (see figure 12.2.2) contains 10 items stacks (each containing 3 items) and 1 order stack plus a conveyer belt (facing the stacks). 12.2.3

The Behavior of Plural-Soar

The canonical simulation of Plural-Soar [32] studied the influence of two cognitive capabilities on organizational performance, that is, the ability to memorize a stack and the ability to communicate. The impact of the number of actors was also studied. Interestingly, increasing the number of actors does not always mean quicker results for the organization. However, more actors does mean less working time per actor. Communication also reduces the working time per actor, provided that there are enough actors with whom to communicate. Waiting time increases when the number


99

item i

item 4

actor X

item 3

A

E

N

item 2

D

C

C

item 1

C

C

F

order stack

stack 1

stack 2

actor Y

stack 10

walkway

conveyer belt

Figure 12.1: Graphical representation of the warehouse task of actors increases, but this effect is reduced by communication and memory skills. Because communication is preferred over search, communicating actors have to wait more than non-communicating ones. Too many actors also reduce the number of answers given, because the waiting time per answer increases. 12.2.4

Critique of Plural-Soar

Plural-Soar, just like Double-AISS, has its drawbacks. Problematic are (a) the cooperation in Plural-Soar, (b) the autonomy of the actors in Plural-Soar and (c) the implementation of the warehouse task: 1. In Plural-Soar, cooperation consists of posing questions about the location of an item and responding truthfully to questions. Whether an actor’s search behavior changes depends on whether it receives an answer (if no answer is received, it engages in exhaustive search). Nevertheless, many of the more elaborate forms of cooperation (such as exchanging tasks) do not occur

100

Simulations of Organizational Problem Solving because agents cannot exchange goals.

2. Actors in Plural-Soar are too autonomous. Each actor can perform the whole warehouse task by itself, so there is no real need for cooperation. A more realistic model would limit the actor’s skills should be limited. 3. Some actions remain implicit in the Plural-Soar model. For instance, the model assumes that an agent perceives the complete content of a stack, once it is in front of it.

12.3

Merging of Double-AISS and Plural-Soar: TASSCS

TASCCS has been built to combine the strong points of both DoubleAISS and Plural-Soar. TASCCS’s starting point was the code of Plural-Soar to which several additions were made (Verhagen, 1992). Added were skills, problem-solving strategies, and organizational structure (implemented as a communication network). Also, evaluation of cooperation requests was mad to depend on the commitment of the receiving actor. However, preferences were absolute, and bounded rationality has not yet been implemented. Also, there are no specific models of other actors to guide communication. TASCCS is an acronym for tasks, actors, structure, commitment, communication, and skills. 12.3.1

The Model of an Actor in TASCSS

As in Double-AISS and Plural-Soar, problem solving is conducted by actors, with every actor being implemented as a separate Soar program. The actors are characterized by their skills, their place in the organizational hierarchy, and their commitment to the organization. As in Double-AISS, a skill is the ability to perform a (sub)task. An


101

actor’s hierarchical position is determined by its place in a communication network. Two communication modes are possible: horizontal and vertical. The actor’s commitment is either altruistic or egotistic. The commitment of an actor determines its preference order with regard to problem-solving strategies and its reaction to received requests. Three problem-solving strategies are available: a) reduce (that is, solve) subtasks, (b) move subtasks to another actor, and (c ) attract subtasks from another actor. Commitment is defined as in Double-AISS: Altruistic actors take the work load of the whole organization into account when deciding on a problem-solving strategy, so altruists prefer “reduce” over “move” and “attract” (because these do not solve (sub)problems, as we know). Egotistic actors take only their own work load into account, and prefer “move” to “reduct” to “attract”. The communication modes and problem solving strategies can be combined into five different alternatives: (a) reduce, (b) a request to move, (c) a command to move, (d) a request to attract, and (e) a command to attract. A request to move can be read as: “Can you do subtask for me concerning item item?” and a command to attract can be read “I want to do subtask for you!” In general, commands are no negotiable, and the only condition under which a command is not obeyed is when it cannot be executed. Requests, in contrast, are evaluated by the receiving actor, according to its preferences. Altruistic actors will accept a move request, but egotistic actors will turn it down. Egotistic actors will accept attract requests (it lowers their work load), and altruistic actors turn them down (they prefer to do things themselves). The ordering of alternatives also depends on the commitment of the actors. Altruistic actors have the following preference order: reduce > attract command > attract request > move request > move command. The egotistic actors have the order: move command > move request > reduce > attract request > attract command. The command mode is chosen when a negative answer is not wanted, and

102


the request mode is chosen when a negative answer is hoped for. Thus, an egotistic actor tries to move a subtask, at first using the command mode, and, if that fails, using the request mode, and turning to less p referred alternatives if that fails too. Altruistic actors prefer reducing an item to attracting an item, and prefer commands over requests, to increase the chance of a successful attraction. “Move” and “attract” apply to subtasks of items. When moving a subtask to another actor, the sending actor decides which subtask is moved. One instance of a subtask is “find”, so an actor may send the message “Can you find item A for me?” When attracting, the sending actor chooses the subtask, but the receiving actor chooses the item. The message being sent has the form: “Can I find an item for you? The warehouse task is split recursively into subtasks (12.2). At the toplevel. the subtasks are take order and fill order. Filling of the order is again split into subtasks, which are finding an item (in the warehouse), and getting the item (getting an item from the stack and putting it on the conveyer belt). In the current model, the skills of the actor are not ordered by preferences, so no skill is preferred over other skills. take order

warehouse task

reduce subtask attract subtask move subtask

Figure 12.2: The warehouse task and its subtasks To implement the warehouse tasks, both “looking at a stack” and “putting an item on the conveyer belt” were made explicit. “Finding an item” consists of moving through the warehouse and examining the stacks. When an item is found, its location is remembered. “Getting


103

hold of an item” consists of going to the item location, manipulating the stack, and taking the item from the stack. This is signaled by the completion marker “in possession”. An item can be put on the conveyer belt by an actor when it is in possession of that actor. The item location has to be communicated when the “get” subtask is moved to (or attracted from) another actor. Also, once the subtask “get” is reduced by another actor, the item has to be handed over to the original subtask owner. The same holds for the moving or attracting the “put” subtask. 12.3.2

The Search Space of an Actor in TASCCS

The search space of an actor in TASCCS consists of several layers. The top layer is the main decision cycle, illustrated in figure 12.3.

empty order list?

no

continue search through problem space

yes orders to take?

yes

take order

end working on file empty?

no

wait

yes end

Figure 12.3: The main cycle of TASSCS First, the actor tries to obtain an order. Then the actor decides whether to reduce the first subtask or to move it. This, of course, depends on the actor’s commitment. When “reduce” is chosen, the skill to perform that subtask is needed in order to be able complete the reduction of the subtask (see figure 12.4, 12.5, and 12.6 respectively).

104


has the actor the find skill?

no

end

yes yes

is the item location known?

remember item location end

move to next stack

item in stack?

no

yes remember item location

end

Figure 12.4: The find subtask

has the actor the get skill?

no

end

yes

is the current llocation the item location?

no

go to the item location

yes

is the wanted item on top?

no

move the top item to another stack

yes take the item from the stack

end

Figure 12.5: The get subtask


has the actor the put skill?

no

105

end

yes

is the item in the actor's possession?

no

go to the location where the item owner is and receive the item

yes put the item on the conveyer belt and remove it form the working on list

end

Figure 12.6: The put subtask When “move” is chosen, the actor to move the subtask to has to be decided on. This is illustrated in figure 12.7. Once the actor to communicate with is known, the information exchange can begin. When all the subtasks have been dealt with, the actor has to decide whether to try to attract a subtask from another actor or take an order. An actor receiving a subtask via a move or attract message has to evaluate that message (see figure 12.8). Once the actor agrees on the request, the item location has to be exchanged or the item has to be handed over, depending on the subtask involved. When this is done, the receiving actor has to go through the same cycle itself. Shall I reduce it or move it? If an actor has reduced a received subtask, the necessary information has to be fed back to the actor who sent the subtask. 12.3.3

The Behavior of TASCCS

Due to limitations in the computer facilities, the organization of the simulation experiment was set to only two actors, with one actor located above the other actor in the organization. The top-level actor is egotistic and the bottom-level actor altruistic. With this organization.

106


is there a message?

end

no

yes is the communication mode command?

yes

is it an attract command? no

no

yes is it an attract request?

no

send yes as an answer

subtask available? no send no as an answer

yes subtask available?

yes

end yes end

no end

evaluate request

send answer

end

Figure 12.7: The move problem space the influence of communication on the problem-solving process of the organization was studied. The dependent variables studied were the amount of cycles waited by an actor, the number of times an actor moves to another stack, and the number of items that is moved to another stack. 12.3.4

Results of TASCSS

The results of TASCSS are summarized in table 1. The results for the same model using Plural-Soar are summarized in table 2. The main difference between our results and the results from Plural-Soar is that the waiting time increases when actors communicate. In Plural-Soar, the waiting time decreases when communi-


107

select communication mode

can actors be reached using this mode?

no

can another mode be used?

no

end

select actor send message

is answer sent?

no

wait

yes no

remove actor from actor list

answer yes? wait

yes

find move subtask

subtask is ...

put

get send item location and move subtask

send own location

has actor arrived?

no

yes give item to actor and move subtask

end end

end

Figure 12.8: The evaluation of a message cation takes place. This is due to the fact that subtasks are solved sequentially in TASCCS. In Plural-Soar, the actor who asks a question waits one cycle for a response. If no response is forthcoming, the agent begins to search for the item. The first actor who asks for an item will search and not wait, as the other actor has not been out to the stacks and so has no information to impart. But the time the first gets his second order and asks for information. the second actor has information that can be imparted. These factors reduce the wait time in Plural-Soar when there are two actors. When there are more than

108


Table 12.1: Summary of the results of TASSCS With Communication Without Communication Actor 1 Actor 2 Σ Actor 1 Actor 2 Σ Actor moves 0 158 158 88 70 158 Item moves 0 50 50 26 19 45 Waited cycles 1621 577 2198 10 62 72 Table 12.2: Summary of the results of Plural-Soar With Communication Without Communication Actor 1 Actor 2 Σ Actor 1 Actor 2 Σ Actor moves 76 82 158 79 81 160 Item moves 20 15 35 17 19 36 Waited cycles 11 18 29 5 32 37 two actors, the actors get in each other’s way, so with communication the wait time increases [32]. Whether the wait time for TASCCS would change for more than two actors is a point for future research. If an answer arrives before the actor has found the item, the speed of problem solving is increased. Waiting in TASCCS is caused both by the communication pattern and the physical properties of the warehouse (i.e., only one actor can look at a stack at one point in time). In Plural-Soar the physical properties of the warehouse, the number of actors, and the actors’ preferences cause the actors to wait. The number of moves of an actor is not influenced by changes in communication and cooperation patterns (of the organization as a whole, that is). The only exception is the Plural-Soar case without communication, where the number of actor’s moves is slightly higher than in the other simulations. The increase of the waiting time in the TASCCS model with communication is in accordance with the theory that well-structured problems need specialization in order to make cooperation useful. Cooperation makes the actors dependent on each other - they have to wait for the other actor to perform a subtask or to give an answer. With


109

specialization the problem-solving process is less tedious. However, specialization also calls for another approach to the communication of subtasks. Subproblems have to be processed in parallel, so that the moving of the “find” subtask allows an actor to take other orders in the meantime. A different implementation of the actions following the completion of the “get” subtask and the moving of the “put” subtask could also speed up the problem-solving process. Not having the original subtask owner as an intermediary item owner but only as a communication manager could be useful, so that after the completion of the “get” subtask, the item doesn’t have to be handed over to the original subtask owner. Instead, the completion can be communicated, after which the communication manager (i.e., the original subtask owner) can send a message, indicating to whom the item should be handed (the actor who accepts the moving of the “put” subtask). Another important improvement would be to let the actors decide when they take an order. Orders can be distributed by an order manager, like the inflow of items in Double-AISS. In this way, the attract strategies not an alternative for the taking of an order, but one of the three possible action strategies.

12.4

Conclusions

TASCCS was designed to overcome some shortcomings of its predecessors Double-AISS and Plural-Soar. The results are in accordance with theoretical predictions about the efficiency of cooperation when solving well-structured problems. To make full use of the benefits of the TASCCS, more work is necessary, however. Future work should focus on the implementation of bounded rationality, the use of revisable belief models of other actors to serve as a guideline for communication, and better evaluation of communication (e.g., based on skills, work load, and preferences, instead of only based on commitment). Labor specialization should be added and subtask reduction done in paral-

110


lel mode to make the advantages of communication and cooperation more clear. Less physical movement and more use of communication will make the problem solving of the organization more effective, from the viewpoint of both physical resources used by actors and a better us of the possibilities communication offers an organization. This can be accomplished by allowing the subtask reduction in parallel mode and by replacing the handing over of items after a moved or attracted “get” subtask to the original subtask owner by the handing over of the item to the actor who reduces the “put” subtask.

Chapter 13 ACTS in action: Sim-ACTS - a simulation model based on ACTS theory Abstract Sim-ACTS is an agent-based computer simulation model of some aspects of ACTS theory [34]. It is the first of a series of models in which increasingly more aspects of ACTS theory will be encompassed. Sim-ACTS studies the effects of communication and organizational structure on organizational problem solving. It is a follow-up study of the TASSCS research project [122]. The simulation experiments from the TASSCS project are replicated using the Sim-ACTS model. The article compares the results of TASSCS’ simulations with those of the Sim-ACTS experiments. Sim-ACTS is implemented in SOAR [74], the implementation of a general cognitive theory [85].

13.1

Introduction

The rise of the agent paradigm within artificial intelligence has given a new impulse to the application of computer simulation in the social sciences. The agent paradigm makes the study of the micro-macro link (i.e., the problem of the relation between the behavior of individuals comprising a social system and the behavior of that system as a whole) more feasible. The use of multiagent simulation models is a natural

112


tool to study the interdependence between individual problem-solving and system behavior. Organization theory has always used models to represent real world situations and to predict what will happen when parameters change. Instead of simulating existing organizations (sometimes called emulation), fictional organizations can be simulated in order to study propositions of an organization theory or to develop a new organization theory. These simulation models are abstracted away from particular organizations, individuals or tasks, but try to use generic concepts related to these three elements. Using agents to simulate individuals seems a natural choice. However, one has to deal with the question of what an agent is or more precisely what type of agent is suited for the particular research goal. The properties of agents can be used to characterize different types of agents as well as the agents’ and system’s possibilities and limitations [33], [123], [124]. The research goal can thus be used to decide on the type of agent and system to be implemented [31], [30]. This article is organized as follows. First, ACTS theory, the theoretical basis of the simulation model, is introduced. Then Soar, the programming language and cognitive theory, is discussed. After this, the simulation model itself is described. Following this is a description of the task that is used to study the interplay between individual properties and behavior and organization properties and behavior. The results of the simulations are then compared to the results of the models predecessor TASSCS in order to validate the model. Finally, some conclusions will be drawn and recommendations for future work will be given.

13.2

ACTS theory

ACTS theory [34] is an extension of the model of bounded rationality [104], [105]. The model of bounded rationality not only specifies that the rationality of human beings is restricted, but also that the

ACTS in Action: Sim-ACTS

113

environment and the relations between the environment and the individual restrict the individuals possibilities. ACTS theory also views agents as limited and constrained by the environment, but extends the model of bounded rationality by specifying the environment and the limitations of the agents’ rationality. In ACTS theory, organizations are viewed as collections of intelligent Agents who are Cognitively restricted, Task oriented and Socially situated. The general concept of bounded rationality is replaced by a full model of a cognitive agent, exhibiting general intelligence. The environment is split into two components, the task and the social situation. Organizational performance and behavior are explained by an interlinked set of computational models. The links are a model of the organizational design. At the micro level, ACTS theory tries to explain how an organizational design affects the behavior and performance of the individual agents communicating and reasoning within a social situation while trying to accomplish a task. At the macro level ACTS theory tries to explain the behavior and performance of groups and organizations with differing organizational designs while the groups or organizations consist of intelligent agents who are socially situated and task oriented. The research questions ACTS theory addresses includes those in which individual decision-making and group decision-making play a key role. What an agent knows and to whom it communicates are important components of ACTS theory. ACTS theory coheres with the ideas in [33] on the nature of the social agent in that the agents’ actions depend on the agents’ cognitive architecture and knowledge. The cognitive architecture does not change over time and is constant over the agents. The information processing, decision making and learning mechanisms are all a function of the cognitive architecture. The agents’ knowledge may change over time, by learning (or perhaps forgetting). The knowledge depends on the agent’s position in the social structure, the task and the problems the organization encompasses. ACTS theory articulates collective organizational constraints and opportunities and how these constraints and opportunities restrict and

114


enable the individual. It thus lets collective phenomena emerge in a dynamic organizational setting. ACTS theory is embodied in a set of fundamental axioms and an expandable set of testable propositions functioning as theorems derived from the axioms. This enables us to implement ACTS theory as a multiagent system that can be used to conduct simulation studies in order to verify its predictions. Earlier work in organization theory can be seen as laying the foundations for ACTS theory. Simon [104] has earlier discussed some issues on the dimensions mentioned in [33], but used a different categorization. Simon argues that some of the (at that time) accepted administrative principles can be on the wrong foot with each other. Instead of giving new proverbs, he argues for more attention on the members of the organization and to the limits on the quantity and quality of the individual’s output. The limits include limits on the ability to perform and limits on the ability to make correct decisions. The limits can be specified as limits that are unconscious (skills, habits, reflexes), limitations on the knowledge needed to perform the task and limits on decision making through the norms and values of the individual. The limits to make correct decisions however, are not analyzed as detailed as in [33].

13.3

Soar

Soar is the computer equivalent of the human problem solving paradigm. Using Soar, one can imitate the behavior of humans solving a particular problem. The underlying structure and the use of operators are all taken from the theory as posed in [86], [85]. Soar, which is derived from the cycle of taking a State, applying an Operator And generating a Result, searches in problem spaces using operators to change the state, problem space or preferences. Problem spaces, operators and states are generated and selected in order to pursue the goals of the system. In order to find correct and efficient paths from the initial state of the system to the desired state, knowledge


115

is needed. This knowledge can be of two forms: directly available as operators or indirectly available through problem resolution. The operators are stored in long-term memory. Problems occur when either more than one or no decision can be made (e.g., two operators can be applied, among which the system then has to choose). The problem resolution becomes the new goal of the system; a technique called subgoaling. When the problem is solved, the knowledge used to solve the subproblem can be stored as a learned production in long-term memory (this is called chunking). Soar’s decision cycle consists of two parts. It first tries to elaborate the current situation with relevant information retrieved from the long-term memory. When all applicable operators have been found, the decision procedure starts. Preferences for operators are processed, and in the end either an impasse is reached or the selected production is fired. The impasse creates a subgoal (as described above). The result of this subgoaling can be stored as a learned operator, with the objects in working memory that caused the impasse as the antecedent and the results of the search in the subgoal problem space in the consequent of the new operator.

13.4

Sim-ACTS

13.4.1

History of Sim-ACTS

Double-AISS [81] was the first completed effort to build an AI-based model of organizational decision making. It was intended as a followup study of Cohen et al.’s the garbage can model [37] of organizational decision making. The garbage can model was also implemented on a numerical computer simulation model. In the Double-AISS model, the influence of various parameters on both individual and organizational level on the problem solving behavior of the individual actors and the organization as a whole was studied. In 1990 Plural Soar [32] was developed at Carnegie Mellon University. This model tries to examine the influence of communication and learning capabilities on the problem solving behavior of independent actors working on a simple

116


task. The TASSCS [122] simulation model was developed to integrate some features of Double-AISS in the Plural SOAR model. Starting from the source code of Plural Soar, a model of communication in organizations was added, reflecting the idea of role taking in DoubleAISS and replacing the broadcasting type of communication in Plural Soar. Added was also the psychological mindset of an agent (egoistic or altruistic) from Double-AISS. Some simulations were carried out and the results were compared to the initial Plural Soar model’s outcomes and theoretical predictions. Sim-ACTS is based on the source code for TASSCS. 13.4.2

The Agents in Sim-ACTS

The agents in Sim-ACTS are implemented in Soar [74]. This implies that they are cognition-based since Soar is based on [85]. Since ACTS theory describes the individual members of the organization as cognitive, the use of Soar is a logical choice. The agents have knowledge on how to solve the task at hand, what parts of the task they can solve (i.e., what skills they have with respect to the task). The agents know that they may be part of a larger organization. They also know with whom they may communicate and how they are placed relative to other agents in the organization structure. The organization structure is implicit in the communication mode that can be used to address other agents. Agents at the same organizational level can be addressed using requests, whereas agents at a lover level in the organization can be commanded. The current implementation does not have a separate communication mode for communicating with agents at higher levels in the organization, however the request mode could also be used for this. The evaluation of received messages is steered by the used communication mode, e.g., an agent receiving a command to carry out some work for another agent will not turn this command down. Requests may be argued about. Agents also have a general psychological mindset, they are either egoistic or altruistic. Their psychological mindset influences their behavior via the prefer-


117

ence ordering among the strategic alternatives that were part of the Double AISS model, i.e., attracting work from another agent, moving work to another agent or reducing the agent’s workload by carrying out the subtask. The ordering of the strategies is determined by the motivation of an agent, is it altruistic, then it is concerned with the workload of the organization as a whole. Is the agent egoistic then it is only concerned with its own workload. So, an altruistic agent has the following ordering of strategies: reduce > attract > move, and an egoistic agent: move > reduce > attract. The agents can be characterized as type III agents in the typology as developed in [123], [124] that is they have norms (implemented as the psychological mindset), goals (the different strategies), an action repertoire (the skills), can be communicated with and react to changes in the environment. The agents however are not value autonomous and currently do not have the ability to build models of other agents. 13.4.3

The Warehouse Task

The warehouse task consists of the filling of a list of orders from a warehouse where the ordered items are stored. The location of the items is unknown to the agents. In search of an ordered item, an agent moves from stack to stack, examining the stacks for the presence of the ordered item (this is the find subtask). When an item is found, it is taken from the stack if it is the top item of that stack. If not, the items on top of the ordered item are moved to another stack in order to free the ordered item after which it can be taken by the agent (this is the get subtask). The ordered item is then placed on a conveyer belt and the agent is finished with the processing of this order (this is the put subtask). In the initial model, all agents are able to perform all subtasks (i.e., finding an item, getting it from the stack and putting it on the conveyerbelt) of which the warehouse task consists.

118

13.5


The Results of the Simulations

The implementation of Sim-ACTS is based on the source code of TASSCS and simulation experiments were carried out to test the results of Sim-ACTS against the outcomes of TASSCS. The simulation experiments consist of the filling of a list of orders from the warehouse. The length and order of the items on the list of orders and the distribution of items over the stacks in the warehouse were not varied. The experiments that were carried out in the TASSCS project were repeated using the Sim-ACTS model in order to be able to compare the results of both models. Sim-ACTS and TASSCS were run in two different setups. In the first setup, the agents were autonomous and had no communication capabilities. The simulation results for Sim-ACTS and TASSCS can be found in table 1 and table 2 respectively. agent X agent Y total organization Agent movements Item movements Orders taken Cycles waited

78 27 8 15

78 16 7 35

156 43 15 50

Table 13.1: Results of the simulation of two independent agents using the Sim-ACTS model. agent X agent Y total organization Agent movements Item movements Orders taken Cycles waited

88 26 8 10

70 19 7 62

158 45 15 72

Table 13.2: Results of the simulation of two independent agents using the TASSCS model. In the second setup, one agent (named agent X) was defined as being able to command the other agent (named agent Y) to perform


119

a specific subtask on a specific item and supplying the agent with all information necessary to perform the subtask. Since agent X was also defined as being egoistic, it preferred to let the altruistic agent Y do all the work (except for the taking of the orders which is not defined as a subtask of the warehouse task in the current implementation). The simulation results for Sim-ACTS and TASSCS can be found in table 3 and table 4 respectively. agent X agent Y total organization Agent movements Item movements Orders taken Cycles waited

0 0 15 1430

312 43 0 525

312 43 15 1955

Table 13.3: Results of the simulation with agent X being egoistic and able to command agent Y who is altruistic using the Sim-ACTS model. agent X agent Y total organization Agent movements Item movements Orders taken Cycles waited

0 0 15 1621

158 50 0 577

158 50 15 2198

Table 13.4: Results of the simulation with agent X being egoistic and able to command agent Y who is altruistic using the TASSCS model.

13.6

Interpretation and Comparison of the Simulation Results

13.6.1

Intra Model Comparison

The number of orders in the warehouse is the same in both simulation models, so the total number of orders taken has not changed. In

120


Sim-ACTS the addition of communication and organizational structure increases the number of agent movements and the number of waiting cycles. In the TASSCS simulations however, the item movements changed and the agent movements were constant. The number of waiting cycles also increased in the TASSCS simulations. The increase in waited cycles when communication and organization structure are added has already been addressed in [122]. After moving a subtask to another agent, the original subtask owner waits for his subordinate to report the accomplishment of the subtask before deciding what to do with the next subtask. This means that roughly speaking only one agent is working at the same time. In the model without communication and organization structure the waiting cycles are caused by physical interdependencies (one agent is processing a stack and the other agent wants to do the same but has to wait for the first agent to finish). The increase in the number of agent movements (Sim-ACTS) or item movements (TASSCS) is due to the fact that instead of two agents moving themselves and orders around in the warehouse in the second setup there is only one agent moving himself and items, thereby changing the physical interdependencies. 13.6.2

Inter Model Comparison

There are some minor differences between the code for TASSCS and Sim-ACTS. One of them is a minor change to the way the moving of items on top of an ordered item is carried out. Therefore instead of an increase in item movements, there is a (slight) increase in the number of agent movements. Another, more important factor impacting the number of agent movements is a different implementation of message handling. Since the new Soar version is more sensitive to retractions of activated production rules, currently the messages are only handled when the agent is not processing any other possible action. So instead of receiving a message at whatever location in the warehouse the receiving agent (i.e., Y) is, the agent only listens to messages when it is at the location where the orders are fetched. This almost doubles


121

the number of agent movements. The number of item movements has decreased in the Sim-ACTS model. This is also due to the minor change in the item moving code. The number of decision cycles spend waiting has also decreased. Both agents wait less. This may be caused by a different working of the Soar decision cycles.

13.7

Conclusions and Discussion

The main difference between TASSCS and Sim-ACTS is the use of the Soar7.x environment instead of the Soar5.x environment. Since Soar5 was implemented in Lisp, it was both slow and craved a lot of memory. Since every agent needed its own Soar process, limitations of access to hardware and networking problems made the use of more than five agents almost impossible. It is therefore that the TASSCS simulations were limited to two agents only. The Soar7 environment is implemented in Tcl, which makes Soar much faster and less memory craving. Within one Tcl process, several Soar agents can be run with even modest hardware constellations. Future simulations can thus examine more interesting organization structures than the one discussed in this work. This will get the simulation model closer to ACTS theory, since Plural Soar, TASSCS and thus Sim-ACTS all lack possibilities to vary the organizational structure sufficiently to investigate its influence on the behavior of individual members of the organization and the organization as a whole. The agents will consequently more closely approximate the Model Social Agent as described in [33]. The current implementation of the receiving of communication is still both highly inefficient and unrealistic. This is one of the major changes proposed for Sim-ACTS-II. Some of the issues already mentioned in the discussion section of [122] are not addressed in the Sim-ACTS model since it is a direct translation of the TASSCS model

122


into the new Soar programming environment. These issues include: • implementation of the attract strategy • implementation of bounded rationality • use of revisable belief models of other actors to serve as a guideline for communication • test the effect of different evaluation strategies of communication (e.g., based on skills, workload and preferences instead of only based on commitment) • test the influence of skill distribution • subtask reduction should be able to run in parallel mode to make the advantages of communication and cooperation more clear and more natural in view of human problem solving • removal of the constraint to hand over an item to the agent that originally owned the subtask This forms the agenda for the development of Sim-ACTS-II. With Sim-ACTS-II, the influence and acquisition of models of other agents will be studied. Generalizing these agent models to role descriptions bring topics such as role switching and inheritance of roles from agents that leave the organization to their replacements within reach. The results of these simulations can then be tested in domain like robotic soccer [73], where agents in real time interact in an ever changing environment and are organized in a soccer team competing with another soccer team. When these agents sufficiently master the basic skills, strategies and roles become the main focus for progress of the level of play. An interesting question is whether an abstract organization theory such as ACTS theory will be able to expand into a down to earth domain as robotic soccer.

Bibliography [1] Rethinking Thinking. The Economist, pages 67–69, December 18th 1999. [2] P.A. Anderson and G.W. Fischer. A Monte Carlo Model of a Garbage Can Decision Process. In J.G. March and R. WeisingerByron, editors, Ambiguity and Command. Pitman, 1986. [3] J.L. Austin. How to Do Things with Words. Oxford University Press, 1962. [4] R.S. Aylett, A.M. Coddington, D.P. Barnes, and R.A GhaneaHercock. What Does a Planner Need to Know About Execution? In S. Steel and R. Alami, editors, Recent Advances in AI Planning, pages 26–38. Springer-Verlag, 1997. [5] K. S. Barber, A. Goel, and C. E. Martin. Dynamic Adaptive Autonomy in Multi-Agent Systems. Journal of Experimental and Theoretical Artificial Intelligence, 2000. To Appear. [6] M. Barbuceanu, T. Gray, and S. Mankowski. How to Make Agents Fulfil Their Obligations. In Nwana and Ndumu, editors, Proc PAAM98, pages 255–276, 1998. [7] J. Baressi and C. Moore. Intentional Relations and Social Understanding. Behavioral and brain sciences, 19(1):107–154, 1996. [8] R.A. Baron. Behavior in Organizations: Understanding and Managing the Human Side of Work. Allyn and Bacon, 1983.

124

References

[9] J. Bates. The Role of Emotion in Believable Agents. Communications of the ACM, 37(7):122–125, 1994. [10] J. Bennett. Rationality. Routledge and Keagan, 1964. [11] M. A. Boden. Autonomy and Artificiality. In M. A. Boden, editor, The Philosophy of Artificial Life, pages 95–108. Oxford University Press, 1996. [12] M. Boman. Norms in Artificial Decision Making. Artificial Intelligence and Law, 7(1):17–35, 1999. [13] M. Boman, J. Andreasen, M. Danielson, C.-G. Jansson, J. Kummeneje, J. Sikström, H. Verhagen, and H. Younes. UBU: Pronouncers in RoboCup teams. In Proceedings of RoboCup Workshop - PRICAI 98, 1998. [14] M. Boman, L. Brouwers, K. Hansson, C.-G. Jansson, J. Kummeneje, and H. Verhagen. Norms for Artificial Agent Action in Markets. In Y. Ye and J. Liu, editors, Proceedings of the IAT99 Workshop: Agents in E-Commerce, pages 255–263, 1999. [15] M. Boman, P. Davidsson, J. Kummeneje, and H. Verhagen. An Anticipatory Multi-Agent Architecture for Socially Rational Action. To be presented at Intelligent Autonomous Systems 6, 2000. [16] M. Boman, P. Davidsson, N. Skarmeas, K. Clark, and R. Gustavsson. Energy Saving and Added Customer Value in Intelligent Buildings. In Nwana, editor, Proceedings PAAM’98, pages 505–517, 1998. [17] M. Boman, P. Davidsson, and H. Verhagen. Pronouncers and Norms. submitted.

References

125

[18] M. Boman, P. Davidsson, and H.L. Younes. Artificial Decision Making Under Uncertainty in Intelligent Buildings. In Proceedings Uncertainty in AI. Morgan Kaufman, 1999. [19] M. Boman and H. Verhagen. Social Intelligence as Norm Adaption. In B. Edmonds and K. Dautenhahn, editors, SAB ’98 Workshop on Socially Situated Intelligence, Centre for Policy Modeling, Technical Report, 1998. [20] A. H. Bond and L. Gasser. An Analysis of Problems and Research in DAI. In A. H. Bond and L. Gasser, editors, Readings in Distributed Artificial Intelligence, pages 3–35. Morgan Kaufmann, 1988. [21] M.E. Bratman. Intentions, Plans, and Practical Reason. Harvard University Press, 1987. [22] M.E. Bratman, D.J. Israel, and M.E. Pollack. Plans and Resource-Bounded Practical Reasoning. Computational Intelligence, 4(4):349–355, 1998. [23] W. Briggs and D. Cook. Flexible Social Laws. In Proceedings of the 1995 International Joint Conferences on Artificial Intelligence, pages 688–693. Morgan Kaufmann, 1995. [24] R. Brooks. A Robust Layered Control System for a Mobile Robot. IEEE Journal of Robotics and Automation, 2(1):14–23, 1986. [25] R. Brown. Group Processes: Dynamics Within and Between Groups. Basil Blackwell, Cambridge, MA., 1990. [26] D. Ca˜ namero, editor. Emotional and Intelligent: The Tangled Knot of Cognition. 1998 AAAI Fall Symposium, AAAI Press, 1998.

126

References

[27] D. Ca˜ namero, C. Numaoka, and P. Petta, editors. Grounding Emotions in Adaptive Systems. Workshop SAB ’98: From Animals to Animats, 1998. [28] K. Carley. Efficiency in a Garbage Can: Implications for Crisis Management. Pitman, 1986. [29] K. M. Carley and L. Gasser. Computational Organization Theory. In G. Weiss, editor, Multiagent Systems—A Modern Approach to Distrivuted Artificial Intelligence, pages 299 – 330. MIT Press, 1999. [30] K.M. Carley. A Comparison of Artificial and Human Organizations. Journal of Economic Behavior and Organization, 31(2):175–191, 1996. [31] K.M. Carley. Artificial Intelligence Within Sociology. Sociological methods and research, 25(1):3–30, 1996. [32] K.M. Carley, J. Kjaer-Hansen, A. Newell, and M. Prietula. Plural-SOAR: Capabilities and Coordination of Multiple Agents. In Warglien M. Masuch, M. and, editor, Artificial Intelligence in Organization and Management Theory. Elsevier Science, 1991. [33] K.M. Carley and A. Newell. The Nature of the Social Agent. Journal of Mathematical Sociology, 19:221–262, 1994. [34] K.M. Carley and M. Prietula. ACTS Theory: Extending the Model of Bounded Rationality. In K.M. Carley and M. Prietula, editors, Computational Organization Theory, pages 55 – 88. Lawrence Erlbaum Associates, 1994. [35] C. Castelfranchi. Multi-Agent Reasoning with Belief Contexts: The Approach and a Case Study. In M. J. Woolridge and N.R. Jennings, editors, Intelligent Agents. Springer-Verlag, 1995.

References

127

[36] C. Castelfranchi. Modelling Social Action for AI Agents. Artificial Intelligence, 103:157 – 182, 1998. [37] M.D. Cohen, J.G. March, and J.P. Olsen. A Garbage Can Model of Organizational Choice. Administrative Science Quarterly, 17:1–25, 1972. [38] J.S. Coleman and T.J. Farraro. Rational Choice Theory: Advocacy and Critique. Sage, 1992. [39] R. Conte and C. Castelfranchi. Cognitive and social action. UCL Press London, 1995. [40] R. Conte, C. Castelfranchi, and F. Dignum. Autonomous NormAcceptance. In Intelligent Agent V: Proceedings of ATAL 98, 1999. [41] R. Conte, R. Falcone, and G. Sartor. Introduction: Agents and Norms: How to Fill the Gap? Artificial Intelligence and Law, pages 1–15, 1999. [42] P. Davidsson. Autonomous Agents and the Concept of Concepts. PhD thesis, Lund University, 1996. [43] T.L. Dean and M. Boddy. An Analysis of Time-Dependent Planning. In Proc Seventh National Conf on Artificial Intelligence, pages 49–54, Minneapolis, Minnesota, 1988. [44] D. Dennett. Brainstorms. Harvester Press, 1981. [45] F. Dignum and H. Weigand. Modelling Communication Between Cooperative Systems. In J. Iivari, K. Lyytinen, and M. Rossi, editors, Proceedings CAiSE95, volume 932 of Lecture Notes in Computer Science, pages 140–153. Springer-Verlag, 1995.

128

References

[46] F. Dignum, H. Weigand, and E. Verharen. Meeting the Deadline: On the Formal Specification of Temporal Deontic Constraints. In Ras and Michalewicz, editors, Proc ISMIS96, volume 1079 of Lecture Notes in Artificial Intelligence, pages 243– 252. Springer-Verlag, 1996. [47] G. Dorais, R.P. Bonasso, D. Kortenkamp, P. Pell, and D. Schreckenghost. Adjustable Autonomy for Human-Centered Autonomous Systems on Mars. Mars Society Conference, 1998. [48] E.H. Durfee and J.S. Rosenschein. Distributed Problem Solving and Multi-Agent Systems: Comparisons and Examples. In Proceedings of the Thirteenth International Distributed Artificial Intelligence Workshop, 1994. [49] L. Ekenberg, M. Boman, and J. Linnerooth-Bayer. General Risk Constraints. Journal of Risk Constraints, 2000. To appear. [50] J.M. Epstein and R. Axtell. Growing Artificial Societies—Social Science from the Bottom Up. MIT Press, 1996. [51] J.D. Erickson. Adjustable Autonomous Operations for Planetary Surface Systems and Vehicles Through Intelligent Systems for Lunar/Mars Missions. In W.W. Guy, editor, NASA-Johnson Space Center (Automation, Robotics, and Simulation Division) FY-96 Annual Report, 1997. Appendix B. [52] M. Fehlin, D. Perlis, M. Pollack, and J. Pollack. Proceedings of the AAAI-95 Fall Symposium Series Rational Agency: Concepts, Theories, Models and Applications. Unpublished, 1995. [53] M.S. Fox. Beyond the Knowledge Level. In L. Kirschberg, editor, Expert Database Systems. Benjamin/Cummings Publishing, 1987. [54] H. Frankfurt. Freedom of the Will and the Concept of a Person. Journal of Philosophy, 1971.

References

129

[55] S. Franklin and A. Graesser. Is It an Agent or Just a Program?: A Taxonomy for Autonomous Agents. In Proceedings of the Third International Workshop on Agent Theories, Architectures and Languages. Springer-Verlag, 1996. [56] L. Gasser. Social Conceptions of Knowledge and Action: DAI Foundations and Open Systems Semantics. Artificial Intelligence, 47:107–138, 1991. [57] M. R. Genesereth and S. Ketchpel. Software Agents. Communications of the ACM, 37(7):48–53, 1994. [58] A. Giddens. The Constitution of Society. University of California Press, 1984. [59] M. Gilbert. On Social Facts. Routledge, 1989. [60] M. Gilbert. Sociality as a Philosophically Significant Category. Journal of social philosophy, 25(3):5–25, 1994. [61] M. Gilbert. Living Together: Rationality, Sociality and Obligation. Rowman and Littlefield, 1996. [62] E. Goffman. The Presentation of Self in Everyday Life. Doubleday, 1959. [63] J. Goldstein. Emergency as a Construct: History and Issues. Emergence, 1(1):49–72, 1999. [64] R. E. Gustavsson. Multi Agent System as Open Societies - a Design Framework. In M.P. Singh, A. Rao, and M.J. Wooldridge, editors, Intelligent Agents IV: Agent Theories, Architectures, and Languages. Springer Verlag, 1998. [65] J. Habermas. The Theory of Communicative Action, Volume One, Reason and the Rationalization of Society. Beacon Press, Boston, 1984. transl McCarthy, orig publ as Theorie des Kommunikativen Handels, 1981.

130

References

[66] B. Hayes-Roth, L. Brownston, and R. Van Gent. Multi-Agent Collaboration in Directed Improvisation. In Proceedings of the First International Conference on Multi-Agent Systems. MIT Press, 1995. [67] B. Hindess. Choice, Rationality, and Social Theory. Unwin Hyman, 1988. [68] N.R. Jennings. On Being Responsible. In E. Werner and Y. Demazeau, editors, Decentralized Artificial Intelligence. Elsevier, 1992. [69] N.R. Jennings. Specification and Implementation of a BeliefeDesire-Joint-Intention Architecture for Collaborative Problem Solving. International Journal of Intelligent Cooperative Information Systems, 2(3):289–318, 1993. [70] N.R. Jennings and J.R. Campos. Towards a Social Level Charaterisation of Socially Responsible Agents. In IEEE Proceedings on Software Engineering, volume 144, pages 11–25, 1997. [71] S. Kalenka and N. R. Jennings. Socially Responsible Decision Making by Autonomous Agents. In Proceedings of the 5th International Colloquim on Cognitive Science, 1997. [72] E. A. Kendall. Agent Roles and Role Models: New Abstractions for Intelligent Agent System Analysis and Design. In A. Holsten, G. Joeris, C. Klauck, M. Klusch, H.-J. M¨ uller, and J. P. M¨ uller, editors, Intelligent Agents in Information and Process Management. Bremen University - TZI, 1998. [73] H. Kitano, M. Asada, Y. Kuniyoshi, I. Noda, and E. Osawa. RoboCup: The Robot World Cup Initiative. In Proceedings of IJCAI-95, Workshop on Entertainment and AI/ALife, 1995.

References

131

[74] J.E. Laird, A. Newell, and P.S. Rosenbloom. SOAR: An Architecture for General Intelligence. Artificial Intelligence, 33:1–64, 1987. [75] K.E. Lange and C.H Lin. Advanced Life Support Program— Requirements Definition and Design Considerations. Technical report, NASA, Lyndon B. Johnson Space Center, Houston, Texas, 1996. Document Number CTSD-ADV-245. [76] I. Leudar. Sociogenesis, Coordination and Mutualism. Journal for the Theory of Social Behaviour, 21(12), 1991. [77] P. Maes. Artificial Life Meets Entertainment: Lifelike Autonomous Agents. Communications of the ACM, 38(11):108– 114, 1995. [78] A.D. Mali. Social Laws for Agent Modeling. In M. Tambe and P. Gmytrasiewicz, editors, Agent Modeling Papers from the AAAI Workshop, pages 53–60. AAAI Press, 1996. [79] J.E. Martin, G.B. Kleindorfer, and W.R. Brashers Jr. The Theory of Bounded Rationality and the Problem of Legitimation. Journal for the Theory of Social Behavior, 17(3):63–82, 1987. ¨ [80] K. Marx. Das Kapital: Kritik der Politischen Okonomie. Meissner, Hamburg, 1867. [81] M. Masuch and P. Lapotin. Beyond Garbage Cans: An AI Model of Organizational Choice. Administrative Science Quarterly, 34:38–67, 1989. [82] J. J. C. Meyer. A Different Approach to Deontic Logic: Deontic Logic Viewed as a Variant of Dynamic Logic. Notre Dame Journal of Formal Logic, 29:109–136, 1988. [83] M. Miceli and C. Castelfranchi. A Cognitive Approach to Values. Journal for the theory of social behavior, 19(2):169–194, 1989.

132

References

[84] M. Mooney Marini. The Role of Models of Purposive Action in Sociology. In J.S. Coleman and T.J. Fararo, editors, Rational Choice Theory: Advocacy and Critique. Sage, 1992. [85] A. Newell. The Knowledge Level. Artificial Intelligence, 18:87– 127, 1982. [86] A. Newell and H.A. Simon. Human Problem Solving. PrenticeHall, 1972. [87] D.A. Norman. Twelve Issues for Cognitive Science. Ablex, 1981. [88] J. W. O’Neill, L. L. Beauvais, and R. W. Scholl. A Structure and Culture Model of Organizational Behavior Variability Reduction. Presented at the Annual Meeting of the Academy of Management 1997. [89] A. Ortony, G. L. Clore, and A. Collins. The Cognitive Structure of Emotions. Cambridge University Press, 1988. [90] S. Ossowski. Co-ordination in Artificial Agent Societies. Springer-Verlag, Berlin, 1999. [91] J.F. Padgett. Managing Garbage Can Hierarchies. Administrative Science Quarterly, 25:583–604, 1980. [92] T. Parsons. The Structure of Social Action. McGraw-Hill, 1937. [93] H. V. D. Parunak. Engineering Artifacts for Multi-Agent Systems. Slides from MAAMAW99 invited presentation to be found at www.erim.org/~van/Presentations/presentations.htm, 1999. [94] B. Reynaud, editor. Les Limites de la Rationalité. Colleque de Cerisy 1993, La Découverte, Paris, 1997. [95] J. Rickel and W.L. Johnson. Animated Agents for Procedural Training in Virtual Reality: Perception, Cognition, and Motor Control. Applied Artificial Intelligence. to appear.

References

133

[96] P. Rizzo, M.V. Veloso, M. Miceli, and A. Cesta. PersonalityDriven Social Behaviors in Believable Agents. In K. Dautenhahn, editor, Proceedings of the AAAI-97 Fall Symposium Series: Socially Intelligent Agents, pages 109–114. AAAI Press, 1997. [97] J. S. Rosenschein and M. R. Genesereth. Deals Among Rational Agents. In Aravind K. Joshi, editor, The Ninth International Joint Conference on Artificial Intelligence, pages 91–99. NorthHolland Publishing, 1985. Also published in The Ecology of Computation, 1998, edited by B.A. Huberman, Elsevier Science Publishers, pp. 117-132, and also in Readings in DAI, 1988, edited by A. H. Bond and L. Gasser, Morgan Kaufmann, pp. 227-234. [98] S.J. Russell and P. Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 1995. [99] N.J. Saam and A. Harrer. Simulating Norms, Social Inequality, and Functional Change in Artificial Societies. Journal of Artificial Societies and Social Simulation, 2(1), 1999. www.soc.surrey.ac.uk/JASSS/2/1/2.html. [100] F. Schick. Having Reasons: An Essay on Rationality and Sociality. Princeton University Press, 1984. [101] J.R. Searle. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press, 1969. [102] Y. Shoham and M. Tennenholtz. On the Synthesis of Useful Social Laws for Artificial Agent Societies (Preliminary Report). In Proceedings of the National Conference on Artificial Intelligence, pages 276–281, San Jose, CA, July 1992. [103] Y. Shoham and M. Tennenholtz. On the Emergence of Social Conventions: modeling, analysis, and simulations. Artificial Intelligence, 94(1-2):139–166, 1997.

134

References

[104] H.A. Simon. The Proverbs of Administration. Public Administration Review, 6:53–67, 1946. [105] H.A. Simon. A Behavioral Model of Organizational Choice. Quarterly Journal of Economics, 69:99–118, 1955. [106] B.A. Skinner. Science and Human Behavior. Macmillan, 1953. [107] R.A. Smit and H.J.E. Verhagen. On Being Social: Degrees of Sociality and Models of Rationality in Relation to Multi-Agent Systems. In M. Fehling, D. Perlis, M. Pollack, and J. Pollack, editors, Proceedings of the AAAI-95 Fall Symposium Series Worjshop: Rational Agency: Concepts, Theories, Models and Applications. Unpublished, 1995. [108] L. Steels. Cooperation Between Distributed Agents Through Self-Organisation. In Y. Demazeau and J.P. M¨ uller, editors, Decentralized AI. Elsevier, 1990. [109] L. Steels. When are Robots Intelligents Autonomous Agents? Robotics and Autonomous Systems, 15:3–9, 1995. [110] P. Stone. Layered Learning in Multi-Agent Systems. PhD thesis, Carnegie Mellon University, Pittsburgh, USA, 1998. CMU-CS98-187. [111] M. Tambe, J. Adibi, Y. Alonaizon, A. Erdem, G. Kaminka, S. Marsella, I. Muslea, and M. Tallis. ISIS: Using an Explicit Model of Teamwork. In RoboCup’97: Proceedings of the first robot world cup competition and conferences. Springer Verlag, 1998. [112] S. E. Toulmin. The Uses of Argument. Cambridge University Press, 1958. [113] R. Tuomela. The Importance of Us: A Philosophical Study of Basic Social Norms. Stanford University Press, 1995.

References

135

[114] A. M. Turing. Computing Machinery and Intelligence. Mind, 59:433–460, 1950. [115] E. Ullman-Margalit. The Emergence of Norms. Clarendon Press, 1977. [116] H. Verhagen. On the Learning of Norms, 1999. MAAMAW’99 poster presentation on CD-Rom. [117] H. Verhagen and M. Boman. Adjustable Autonomy, Norms and Pronouncers. In D. Musliner, B. Pell, G. Dorais, D. Kortenkanp, N. Muscettola, and M. Tambe, editors, Proceedings of the AAAI 1999 Spring Symposium Workshop: Agents with Adjustable Autonomy, pages 124–127. AAAI Press, 1999. [118] H. Verhagen and M. Boman. Norms Can Replace Plans. In D. Kortenkamp, editor, IJCAI99 Workshop PLAN-1: Adjustable Autonomy Systems, pages 65–67, 1999. Invited paper. [119] H. Verhagen and J Kummeneje. Adjustable Autonomy, Delegation, and Distribution of Decision Making. In Proceedings CEEMAS ’99, pages 301–306, 1999. [120] H. J. E. Verhagen. ACTS in Action. In J. A. Sichman, R. Conte, and N. Gilbert, editors, Mulit-Agent Systems and Agent-Based Simulations, number 1534 in Lecture Notes in Artificial Intelligence. Springer, 1998. [121] H.J.E. Verhagen and M. Boman. Norm Spreading as Social Learning. in preparation. [122] H.J.E. Verhagen and M. Masuch. TASCCS: A Synthesis of Double-AISS and Plural-SOAR. In K. Carley and M. Prietula, editors, Computational Organization Theory. Lawrence Erlbaum Associates, 1994.

136

References

[123] H.J.E. Verhagen and R.A. Smit. Modelling Social Agents in a Multiagent World. In W. van de Velde and J.W. Perram, editors, Position papers MAAMAW 1996, Technical report 961. Vrije Universiteit Brussel - Artificial Intelligence Laboratory, 1996. [124] H.J.E. Verhagen and R.A. Smit. Multiagent Systems as Simulation Tools for Social Theory Testing. Paper presented at poster session at ICCS and SS Siena, 1997. [125] J. M. Vidal and E. H. Durfee. The Impact of Nested Agent Models in an Information Economy. In M. Tokoro, editor, Proceedings of the Second International Conference on Multiagent Systems, pages 377–384, Menlo Park CA, 1996. AAAI Press. [126] G. H. von Wright. An Essay in Modal Logic. North-Holland, 1951. [127] M. Weber. Wirtschaft und Gesellschaft. T¨ ubingen, 1922. [128] E. Werner. Logical Foundations of Distributed Artificial Intelligence. In G.M.P. O’Hare and N.R. Jennings, editors, Foundations of Distributed Artificial Intelligence. Wiley, 1996. [129] M. Wooldridge and N.R. Jennings. Intelligent Agents: Theory and Practice. Knowledge Engineering Review, 10(2):115–152, 1995. [130] M. J. Wooldridge and N.R. Jennings. Agent Theories, Architectures and Languages: A Survey. In M. J. Woolridge and N.R. Jennings, editors, Intelligent Agents:, ECAI Workshop on Agent Theories, Architectures and Languages. Springer-Verlag, 1995. [131] H. Younes. Current Tools for Assisting Real-Time Decision Making Agents. Master’s thesis, DSV, Royal Institute of Technology, Stockholm, Sweden, 1998. no. 98-x-073.