unpredictable domain of forest fire, the one realated to the individuation and integration of planning techniques suitable to this ..... Trial by fire: Understanding the ...
Planning in a Complex Real Domain1 F. Ricci, A. Perini and P. Avesani, Istituto per la Ricerca Scientifica e Tecnologica 38050 Povo (TN) Italy e.mail: {ricci,perini,avesani}@irst.it
Abstract Dimensions of complexity raised during the definition of a system aimed at supporting the planning of initial attack to forest fires are presented and discussed. The complexity deriving from the highly dynamic and unpredictable domain of forest fire, the one realated to the individuation and integration of planning techniques suitable to this domain, the complexity of addressing the problem of taking into account the role of the user to be supported by the system and finally the complexity of an architecture able to integrate different subsystems. In particular we focus on the severe constraints to the definition of a planning approach posed by the fire fighting domain, constraints which cannot be satisfied completely by any of the current planning paradigms. We propose an approach based on the integratation of skeletal planning and case based reasoning techniques with constraint reasoning. More specifically temporal constraints are used in two steps of the planning process: plan fitting and adaptation, and resource scheduling. Work on the development of the system software architecture with a OOD methodology is in progress.
Introduction In this paper is described the current state of development of a system supporting the planning of initial attack to forest fires. This system is part of a decision support system aimed at supporting the user in the whole process of the forest fire management including both situation assessment and planning activities. One of the goal of this paper is to present and discuss some topics related to AI planning that we have rediscovered in studying the fire fighting domain. In fact our research starts from a deep analysis of the domain that we have pursued collecting information from many sources: manuals, books, regional laws and plans, interviews with firemen and responsible of anti-fire centres. The development of the system is still in an early phase. We have completed a first cycle of knowledge acquisition interviews and we have defined a complete functional architecture. We are addressing now the software architecture of the system. This architecture will be developed using the design methodology proposed by G. Booch [Booch 91] and with the aid of a software tools OSD, that implements faithfully the Booch methodology. In a classical AI perspective there is a neat distinction of roles between the user and the planner: the user poses the goal, the environment sets the initial conditions and the planner finds out the solution. This simplified view of the problem solving activity does not apply in our case. The user is an expert of the domain and want to control the global information flow. In some cases he is able to solve the current goal, for example mostly regarding strategical decisions, in some other cases he want to set constraint to the search process. Decisions are always taken by the user, using the computer together with other tools and usually following a complex operational flow. This discussion points out the requirement to build a system that can be used in an unstructured way. Assumption may be done in any moment and change of focus of reasoning are continuous, for example new fire-alarms may arrive at any moment of the planning process. In our opinion these aspect have not been addressed properly by the AI research on planning that is mainly motivated by the desire to build really intelligent agents, capable of autonomous behaviour in complex environments. So, many of the great difficulties of AI planning systems, have only a limited interest for this research. Beside, our main difficulty is to enable, with an effective problem solving architecture, a smooth co-operation between the human agent and the planner. This yields another dimension of complexity: the architectural problems. The necessity to couple strictly the user interaction with the functionalities of the planner subsystem sets complex architectural problems. First of all, we must mention the integration of different reasoning paradigms, that is an extensively studied problem in AI [Skinner and Luger 92] [Wilkins and Desimone 92] [Kambhampati et al 91] [Hayes-Roth 85]. Second, the planning system is only a component of a bigger integrated system. The planning component communicates with another component devoted to situation assessment that is in charge of acquiring and processing all the data 1 This work has been partially supported by the Esprit III project CHARADE (#6095).
coming from the environment and from the other information sources (data bases and knowledge bases). This interaction is controlled by the Man Machine Interface that, based on an explicit representation of the user tasks controls the dialogue. Therefore we have decided to use a specific submodule of the MMI as general controller of both the system and the problem solving process. The planner in this view is passive and is activated in each step and functions by explicit user requests, that are controlled by the task model. This approach will be described more extensively on a future work. Many computerised systems have been proposed and developed to help the responsible organisations in dealing with the management of environmental emergencies. These application are usually based on traditional techniques (GIS, simulation models, resources management with data base) very few AI rooted applications have been developed. In the domain of fire fighting we cannot avoid quoting the Phoenix project [Cohen et al. 89], which is real-time adaptive planner that manages forest fires in a simulated environment, and the system developed by P. Kourtz that addresses the problem of dispatching waterbombers, helicopters and crews for fire control in Quebec [Kourtz 87]. The first one is a research on the design of intelligent agents and the second is a classification system based on rules implemented in PROLOG. In the following sections we shall discuss first the severe constraints arising from the forest fire domain. Second, we shall propose our provisional architecture and finally we shall address some open problem and we shall indicate our future directions of research.
The forest fire domain The complexity of the forest fire domain derives from features which are typical of environmental problems. Fire is a dynamic phenomenon whose evolution is determined by weather conditions, in particular wind intensity and direction, by humidity, by fuel type, that is, by parameters which usually change rapidly and sometimes in an unpredictable way. In this context the temporal dimension of decision making and of the operations for fire fighting is of fundamental importance. Operational constraints impose often quick decisions that drastically limit the possibility to build a plan. Moreover relevant fire events can happen on very different time and spatial scales, from seconds to days, from meters to kilometers, determining a large variety of world states. Data are always incomplete and uncertain, unless, in some cases, totally absent. The role of past knowledge is extremely important, sub optimal solutions are often adopted, because of the bias in the decision process. The management of forest fire, as in general environmental emergency, involves organisations which have decisional and operative centers distributed on the territory. So managing a forest fire can require several centers to cooperate. Moreover complex coordination problems can araise when resources from different organisations, like forestry people, police, ambulances are employed. These are all general features of the fire domain. To be more specific, the management of forest fires follows a precise operational workflow that is typical of each fire fighting organisation. We shall concentrate on the work organisation in an Italian Provincial center. The user of the system is the controller based in a provincial center. His tools are: a workstation, a dedicated line to acquire data from infrared sensors and meteo sensors, a radio, a fax, a telephone and a printer. The system running on the workstation comprises a Geographic information System, a graphical simulator of the fire evolution, tools for territorial, meteo and resource assessment and a module for supporting the intervention planning and control. When a new fire is reported, the alarm is promptly validated and the situation assessed by the user possibly running the fire propagation model. Alternating phases of planning and situation assessment he will define the intervention plan, choose and dispatch the necessary resources. The control of the intervention passes, now, to field controllers. In this context we can view the planning problem as consisting in: • Formulating a goal. For example extinguishing the fire, retarding the fire, arresting the fire, evacuating the zone, reporting about a situation. • Choosing the intervention strategy.For instance performing fire recognition, attacking the fire simultaneously on different sectors of the fire front ("sectorization"), or combining basic strategies. • Finding tactics that allow a correct strategy implementation. A tactic is a partially ordered set of actions implementing specific fire fighting techniques. Typical actions are the construction of a fire line, the transportation of means, the recognition activity of a scout. The correctness of a tactic refers to the quantity of water, the resources to be employed and the temporal deadline associated to the corresponding strategy.
• Assigning the resources. The actions require certain type and quantity of resources. Resources have to be taken from bases or other fires and allocated to actions taking into account time constraints of the plan itself and constraints on resource allocation periods.
The intervention planning approach Dealing with the design of the planning system we have at once faced the necessity to integrate different planning techniques. No planning paradigm offers a complete solution to a real practical planning problem, as already discussed in the introduction. Therefore we have excluded the possibility to implement faithfully a general methodology. Based on a reasonable level of description of a plan we have analyzed planning techniques that appear most adequate to the specific feature of our planning problem. We report here some basic conclusions of this analysis. Main limitations for applying a constructive planning approach to the fire fighting domain rest on the inadequacy of the action representation and on the assumption that the world and the consequences of actions may be completely predicted (see for instance [Wilkins 84]). Skeletal planning [Friedland and Iwasaki 85] and Case Based planning [Hammond 89] assigne a foundamental role to the past experience in formulating new plans and in debuggigng old ones. In both cases plans are fetched from the memory, but they radically differ in the meaning given to a stored plan. In Case Based planning a plan represents a single specific experience, while a skeletal plan is a generalization, done by the expert, of many similar experiences. There are three great advantages, with these approaches which meet our application requirements: • plan reuse which allows to save computation time, a delicate issue in interactive systems; • cognitive adequacy, that is taking into account the fact that in some domains relatively little time is spent in reasoning from starting state to goals or backwards from goals to starting states. In the fire fighting domain the expert has no general theory to face the planning problem and he doesn't invent a figth plan from scratch. Conversely the expert has a repertoire of general strategies and an evolving range of techniques. • no necessity of a theory of world evolution. The integration of specialized constraint based techniques for representing and reasoning about temporal information have been used in planning approaches for dealing with "realistic problems" where actions take time and can interact when performed simultaneously or in scheduling systems aimed at solving finite resource scheduling problems (see, for instance [Dean et al. 1988], [Allen 1991]). Planning for forest fire management faces with similar problems. This suggests to take advantage of available temporal reasoning techniques to represent and reason about temporal information on the actions of an intervention plan and on the resources to be allocated to those actions. Interesting examples of integrations of the mentioned techniques can be found in the literature. For instance the integration of skeletal and Case Based reasoning has been proposed in [Alexander and Tsatsoulis 91], the integration of Case Based reasoning and constraint reasoning in [Hinrichs 88].
Dynamic Data Space Fire Data Action Net
Plan Reasoner
Past Interventions
Resource Data
Constraint Reasoner
Action
Constraint
Static Data Space
Figure 1: The intervention planning system architecture
A basic point to be stressed again is that technical choices, in our project, have been driven by domain requirements and constraints. So, for instance, the integration of CBR techniques that can provide quick and adequate suggestions to the user with a constraint reasoner that manages automatically the consistency of the information available to the user faces the requirement to produce a system for supporting the user during planning activities. Also the choice to use both skeletal and CBR techniques takes into account results from knowledge acquisition from the fire fighting domain which report on plan reasoning two different levels: the strategical level which consists in making decisions for the global intervention and the tactical level which refers to a local view of the intervention, for instance dealing with a specific fire front sector.We represent intervention strategies as skeletal plans which are classified respect to the type of environmental scenario they are suitable for. They correspond to the consolidate component of the past interventions memory. While instances of tactics applied in previous interventions are stored into a case memory. We summarize our approach by briefly commenting Figure 1 which depicts the architecture of the intervention planning system. Three main areas are highlighted: • the static data space which contains domain knowledge such as past intervention strategies and tactics, specific fire fighting techniques (Action Library) and domain constraints; • the dynamic data space containing fire data produced by the assessment of a fire alarm, the temporal dimention of tactics retrieved for the current situation (called Action Net), static and dynamic information on resources; • two reasoning modules: the plan reasoner and the constraint reasoner. The main functionalities of the planning system can be described following the fire fighting problem definition stated in the previous section. The planning problem associated to a given fire alarm is described in terms of the emergency scenario which is defined by a set of fire parameters like fire location, meteo conditions, topography, available resources that are located in the bases close to the fire (see Fire Data) and a set of appropriate goal such as extinguishing the fire, evacuating a zone or performing a recognition within time deadlines. Finding a strategy means to find an association between the description of the current scenario and one of the preclassified possible strategies. The refinement of a selected strategy into appropriate tactics is performed through CBR techniques. The retrieval of tactis is performed by a partial matching algorithm that looks for goals similar to the one associated to the strategical step to be expanded and to scenario descriptions similar to the current one. A tactic contains the temporal dimension of a plan through information such as action duration, possible time constraints respect to the starting and ending times of the plan, and temporal relation between actions. These information are expressed as a set of temporal constraints defined on the temporal variables, such as the starting and ending times of actions. They are represented through constraint network (Action Net) such as the one depicted in Figure 2. Temporal constraint can represent qualitative temporal information, such as the fire starting time is before the fire detection time, or the action for setting up a reservoir ends before or at the beginning of the action in which the reservoir is used, or metric information. Example of metric constraints are the action duration constraints. For instance the installation of an artificial water supply for the helicopter, in the plan depicted in Figure 2, can take between half an hour and one hour and fifteen minutes and is represented as: 0:30 ≤ tend - tstart ≤ 1:15 hours.
Helicopter Recognition
