A Framework for Distributed Problem Solving

Situations in which several agents must interact to achieve goals present difficulties of coordination and cooperation not found in single-agent problem solving contexts. Techniques for coordination and cooperation required in group problem solving are not well understood because most AI models deal with cases in which problems are solved by a single agent. In this paper we present a framework for distributed problem solving that describes some of the expertise an agent working in an multi-agent environment must have. An application of the framework to the domain of air-traffic control is discussed. Here each aircraft is viewed as an agent that must cooperate with others to achieve a conflict-free plan. 1. The agent must be able to continuously update possibly incomplete or incorrect world models (caused by limited situation assessment or by the existence of agents that change the environment dynamically and unpredictably). 2. The agent must be able to integrate information coming asynchronously from other agents. 3. The agent must know when new information invalidates his current attempt to plan and be able to modify his plan during execution if new and inconsistent information is uncovered. 4. The agent must be able to interrupt his planning to help others acquire knowledge to achieve their goals (especially when his goals and theirs interact or conflict). 5. The agent must be able to get others satisfy his goal or play his role. to help Not all these competences are required of single-agent problem solvers, and, since most existing models of problem solving (e.g., STRIPS Dl, BUILD [Z], NOAH [3] and many others) focus on single-agent problems, little is known about how agents solving distributed problems achieve such competences. We have been attempting to understand distributed problem solving by developing computational models of agents that have these abilities. In this paper we present a view of agents in group problem solving situations that we have developed, and discuss a specific problem solver for distributed air-traffic control (DATC) consistent with this view. II A VIEW OF DISTRIBUTED PROBLEM SOLVING ---Using competences such as those discussed above, we have evolved a view of the important features of distributed problem solving agents that differs considerably from the view traditionally adopted in models of single-agent problem solvers. For example, many theories [l], [Z] suggest that a central feature of the single-agent problem solver, is that its activities are decomposed into separate, strictly ordered, phases of information gathering, planning, and execution. However, the above competences indicate that, in situations where a given agent is not the sole cause of change, and therefore where not all important consequences of a planned action can be foreseen at the time of planning, it is essential that the agent be able to effectively interweave information gathering, planning and execution tasks.