Department of Computer Science and Business Administration ... The agent metaphor has been used successfully in computer program development and theory for several years now, and the number of agent applications is increasing rapidly ...
Test Implementations of Information and Decision Support Systems
Paul Davidsson Department of Computer Science and Business Administration University of Karlskrona/Ronneby Soft Center, S-372 25 Ronneby, Sweden
Magnus Boman Department of Computer and Systems Sciences Stockholm University and the Royal Institute of Technology Electrum 230, S-164 40 Kista, Sweden
1 Introduction The agent metaphor has been used successfully in computer program development and theory for several years now, and the number of agent applications is increasing rapidly [5]. Roughly speaking, by agent we mean an independent piece of software often implemented as a separate process capable of communicating with other agents, and typically having well-defined internal state, behavior, and goals. The concept of software agent, as well as the communication languages used by such agents to exchanges messages, is becoming a standard concept in software engineering [6]. Of special interest to developers of real-life applications with computers acting as monitors, guards, or controllers of several autonomous processes or functions, is the current upsurge in the research and development area of multi-agent systems (MAS).
The MAS area is concerned with the interaction between agents and the environment in which they are executed. This environment is often interesting enough to merit some effort being put into its design, in addition to that of the individual agents in the MAS. For instance, it is in many system necessary to have global constraints for local agent behavior. The characteristics of such constraint are often put in anthropomorphic terms, and agents are as a result described as, e.g., selfish but socially aware entities. In this paper, we are interested in the advantages of adapting a MAS design policy for the inherently distributed domain of DA/DSM and in particular value-added services in the context of intelligent buildings.
We will describe a test implementation of an intelligent building MAS application. For a related test implementation, we refer to the paper by Hans Akkermans and Fredrik Ygge in this proceedings. Both test implementations make use of the existing power lines in the building for communication between agents. We are currently experimenting with simulations, and later this year fielded experiments in Villa Wega (Ronneby, Sweden) will commence. A large part of the required hardware is already in place in Villa Wega.
2 Energy Saving and Enhancement of Customer Value: An intelligent building application The system consists of a collection of software agents that monitor and control an office building using the electrical devices already present in the building. The objectives of the application are both energy saving and enhancement of customer value through value-added services. Energy saving in this environment is realized by controlling lights, heating, ventilation, etc. Examples include lights that are automatically switched off, and the room temperature being lowered in an empty room.
Enhancement of customer value is realized by the system by taking into account the needs and wants of the people in the building. For instance, by adapting temperature and light intensity according to each person’s personal preferences.
The goal is to make the system transparent to the people in the building in the sense that they do not have to interact with the system in any laborious manner. The system automatically detects in which room in the building each person is at any moment and adapts the conditions in the room according to that person’s preferences. (Of course, when the preferences are entered into the system initially or changed during execution, some more sophisticated interaction is necessary.)
2.1 The Simulation Even though the goal is to implement a fielded application in Villa Wega, a first step is to simulate this application in order to evaluate the viability of the approach. The software used in these simulations consisted mainly of two separate components: (1) A society of agents, i.e., the MAS, that communicate and cooperate in order to achieve their goals, i.e., energy saving and enhanced customer value. This part was mainly implemented in the April language and its extensions [2]. For more information about this component we refer to [4]. (2) The simulation environment, which simulates the building and its electrical devices and the behavior of real persons present in the building. It provides input to the agents in the MAS and react to output received from the agents in the MAS. In addition, it has a graphical user interface that visualizes the state of the building. This component is implemented in the Java language. For more information about this component we refer to [1].
In addition, there is a communication interface between the components (described in [1]). We thus have the following components:
simulation environment
interface
MAS
Our intention is to use the same MAS in the fielded application. It will be interfaced to the actual environment, i.e., Villa Vega, which will substitute the simulation environment. We will then have the following situation:
Villa Wega
interface
MAS
2.2 Available Electrical Devices A building contains a number of electrical devices that constitute an important part of the infrastructure of the building. The general idea behind the application is that these devices will be interacting with, and be controlled by, the MAS. They will provide input to the MAS and occasionally receive instructions from it. Some of the devices are sensory and some are actuator devices.
The sensory devices (read only) assumed are: temperature, light intensity, fire detector, presence (detects whether there is activity in a room in a room or not), and active badges. The active badge system makes it possible for the MAS to know which persons are in each room at any moment. A vision for future work is to let the MAS know in which room each person will be in at, for example, one hour (e.g., by having access to each person’s electronic time manager).
The actuator devices differs from the sensory devices in that it is possible, besides reading the state of the device, to change the state of the device (in order to change the state of the building). The actuator devices in the current application are lamps, radiators, and generic mobile devices that can be connected to an arbitrary electrical device, e.g., a coffee machine or a personal computer. It is possible switch on and off the device connected to the generic mobile device and to read its state.
For the communication between agents and devices LONWORKS network technology is used. More details about this communication and about the devices mentioned above can be found in [1].
2.3 Types of Agents There are four main categories of agents in this application:
• Environmental Parameter (EP) agents, which monitor and control environmental parameters in a particular room by having access to sensor and actuator devices for reading and changing the environmental parameter. For instance, a temperature agent can read the temperature sensor and control the radiators in a room. • Badge agents, which keep track of where in the building (in which room) each person (badge) is situated at any time. • Room agents, which each corresponds and controls a particular room with the goal of saving as much energy as possible, e.g., by controlling the radiators and lights of the room. • Personal Comfort (PC) agents, which each corresponds to a particular person. It contains personal preferences and acts as a surrogate of the person in the multi-agent system trying to maximize customer value. Thus, the agent does not model the behavior of a person, rather it tries to act on that person’s behalf, in his/her interest. We see that the goals of the room agents and the personal comfort agents may be in conflict: the room agents maximizing energy saving and the personal comfort agents maximizing customer value. The problems caused by conflicting goals can be solved either through negotiation between the involved agents or by consulting a decision module [7].
Moreover, in order to deal with changes in the infrastructure of the building and in the personnel, the system can be re-configured dynamically. It is possible to add new agents at run-time without the need for the interruption of the normal operation of the system.
2.4 System Constraints The system conforms to a number of general rules (constraints) that can be considered global to the application. Some of them are: • If and only if at least one person is in the building, the corridor light must be on. • When a particular person is in the building his/her office (and other relevant rooms) must adapt temperature, light, etc. to his/hers preferences, otherwise the corresponding electrical equipment should be turned off (or similar). • It must always be possible to over-rule the decisions of the agents in the MAS by physical interaction with the electrical equipment. For instance, even if an EP agent has decided that the light in a room should be on, it must be possible for a person to turn off the light using the switch in the actual room.
2.5 Interaction between Agents Interaction between agents is mainly achieved through message passing adopting a KQML [3] style for the format of the messages accepted. This is due to the wide acceptance of KQML in the agent community.
This is a very simple example of a sequence of messages that are sent between a small set of agents: • Agent-John, which is the personal comfort agent of the real person John, the preferences are: temperature = 21 C and light = normal, • Agent-Corridor and Agent-RoomJohn, which are room agents, • Agent-Badge, which is the badge agent, • EP-LightCorridor, EP-LightRoomJohn, EP-TemperatureCorridor, and EP-TemperatureRoomJohn, which monitor and control corresponding environmental parameters. When John enters the building, Agent-Badge detects this and sends a message telling about this to Agent-John and Agent-Corridor. Agent-Corridor sends (if nobody already is in the building) a message to EP-LightCorridor to turn on the light. Agent-John sends message to Agent-RoomJohn about John’s preferences concerning temperature and light. Agent-RoomJohn then sends messages about this to EP-TemperatureRoomJohn and EP-LightRoomJohn. When John exits the building, AgentBadge sends a message about this to Agent-John and Agent-Corridor. Agent-John then tell AgentRoomJohn that John no longer is in the building, who, in turn, notifies EP-TemperatureRoomJohn and EP-LightRoomJohn to no longer pay attention to John’s preferences. Similarly, Agent-Corridor EP-LightCorridor sends (if nobody is left in the building) a message to EP-LightCorridor to turn off the light.
In a more advanced version, Agent-John has access to John’s electronic time-manager. It could then in advance (e.g., an hour before John is scheduled to enter the building) send a message to AgentRoomJohn saying that the temperature should be set to 21 C in the most economic way possible.
3 Conclusions and Future Work We have sketched a current project aimed at investigating the usefulness of utilising the agent metaphor and the notion of multi-agent systems for the design of control systems for intelligent buildings. The inherently distributed domain is well-suited for such trials, and the simulation setting described
is mapped to a physical building with hardware implementations already made. The next step will be to evaluate the simulations, before the upcoming transition from simulation testbed to physical implementation. Moreover, we believe that applications as the one described above which combines energy saving and enhancement of customer value will play an important role in supporting customer retention.
Acknowledgements The authors would like to thank Martin Fredriksson and Daniel Ljungberg (M.Sc. students at the University of Karlskrona/Ronneby) who implemented the simulation environment, Nikolaos Skarmeas and Professor Keith L. Clark (consultants at Agent Solutions Partnership, UK) who implemented the MAS, and the former project leader Staffan Hägg.
References [1] Boman M., Davidsson P., Fredriksson M., Hägg S., Ljungberg D., and Skarmeas N., Specification of a Simulation Environment for Energy Saving and Customer Value Applications in Home Automation, Technical Report, Ronneby Sweden, 1997. [2] McCabe F. G. and Clark K. L., April: Agent Process Interaction Language, in Wooldridge M. J. and Jennings N. R. (eds.), Intelligent Agents, Lecture Notes in Artificial Intelligence, 890, pp. 324-340, Springer-Verlag, 1995. [3] Finin T., Fritzson R., and McKay D., et al., An Overview of KQML: A Knowledge Query and Manipulation Language, technical report, Department of Computer Science, University of Maryland, Baltimore County, USA, 1992. [4] Boman M., Davidsson P., Fredriksson M., Hägg S., Ljungberg D., and Skarmeas N., Energy saving and customer value in home automation: a basic multi-agent system application specification, Technical Report, Ronneby Sweden, 1997. [5] Jennings N.R. and Wooldridge M., Applying Agent Technology, Applied AI, Vol. 9, No. 4, pp. 357-369, 1995. [6] Genesereth M.R. and Ketchpel, Software Agents, Communications of the ACM, Vol. 37, No. 7, pp. 48-53, 1994. [7] Ekenberg L., Boman M., and Danielson M., A Tool for Coordinating Autonomous Agents with Conflicting Goals, First International Conference on Multi-Agent Systems, pp. 89-93, 1995.
BIOGRAPHICAL INFORMATION Authors:
Paul Davidsson and Magnus Boman
Company:
University of Karlskrona/Ronneby and Stockholm University
Country:
Sweden
Paul Davidsson is assistant professor at the Department of Computer Science and Business Administration, University of Karlskrona/Ronneby, Sweden. He received his Ph.D in Computer Science in 1996 from Lund University, Sweden. His research interests includes the theory and application of multi-agent systems, autonomous agents and machine learning. He is project leader for Distributed Decision Islands, which is a sub-project within the ISES project. The ISES project is financed and managed by EnerSearch AB.
Magnus Boman leads the research group DECIDE at the Department of Computer & Systems Sciences (DSV), Stockholm University and the Royal Institute of Technology, where he has a permanent tenure position. He graduated at DSV in 1993 and has since pursued his interests in decision making agents (human and artificial), as reported in several journal articles and invited talks.