Multiagent classical planning

Abstract

Classical planning problems consist of an environment in a predefined state; a set of deterministic actions that, under certain conditions, change the state of the environment; and a set of goal conditions. A solution to a classical planning problem is a sequence of actions that leads from the initial state to a state satisfying the problem’s goal conditions. There are many methods for finding solutions to classical planning problems, and a popular technique is to exploit structures that commonly occur. One such structure, apparent in many planning domains, is a breakdown of the problem into multiple agents. However, methods for finding and exploiting multiagent structures are not prevalent in the literature and are currently not competitive. This thesis sets out to rectify this problem. Its first main contribution, is to introduce a domain independent algorithm for extracting multiagent structure from classical planning problems. The algorithm relies on identifying a generalisable property of agents in planning; namely, that agents are entities with an internal state, a part of the planning problem that, under a certain distribution of actions, only they can modify. Once this is appropriately formalised, the decomposition algorithm is introduced and is shown to produce identifiably multiagent decompositions over all of the classical planning domains used in the International Planning Competitions, even finding more detailed decompositions than are used by humans in certain cases. Solving multiagent planning problems can be challenging because a solution may require complex inter-agent coordination. The second main contribution of the thesis is a heuristic planning algorithm that effectively exploits the structure of decomposed domains. The algorithm transforms the coordination problem into a process of subgoal generation that can be solved efficiently under a well-known relaxation in planning. The generated subgoals guide the search so that it is always performed by one single-agent subproblem at a time. The algorithm is evaluated and shown to greatly outperform current state-of-the-art planners over decomposable domains. The thesis also includes discussion of the possible extensions of this work, to include the multiagent concepts of self-interested agents and concurrent actions. Results from the multiagent planning literature are improved upon and a new solution concept is presented that accounts for the ‘farsightedness’ inherent in planning. A method is then presented that can find stable solutions for a certain class of multiagent planning problems. A new method is introduced for modelling concurrent actions that allows them to be written without requiring knowledge of each other agent in the domain, and it is shown how such problems can be solved by a translation to single-agent planning.