Abstract: Model checking of temporal logic has already been proposed for automatic planning. In this paper, we introduce a simple adaptation of the ATL model checking algorithm that returns a strategy to achieve given goal. We point out that the algorithm generalizes minimaxing, and that ATL models generalize traditional game trees. The paper ends with suggestions about other game theory concepts that can be transfered to ATL-based planning.

Abstract: The design of complex systems needs sophisticated coordination mechanisms. This paper proposes an approach based on multi-agent planning to coordinate such systems in the tactical aircraft simulation domain. This paper presents an integral cycle, from modelling to validation, that enables the building of feasible multi-agent plans as developped in a project named SCALA promoted at Dassault Aviation.

Abstract: Many tasks require a team of agents to act together in a coordinated way in a complex, uncertain environment. Examples include search and rescue, control of a complex system such as a factory, or robot soccer. Such tasks involve many agents, and huge numbers of states and possible actions. Factored Markov decision processes (MDPs) provide a formal foundation for modeling such complex systems naturally and compactly. We propose a framework for multi-agent coordination and planning in factored MDPs based on the use of a factored value function — an approximate decomposition of the team value function as a sum of value functions of small subteams of agents. We show that factored value functions naturally give rise to an optimal distributed algorithm for joint action selection, whose communication structure naturally mirrors the interactions between the subteams. We present an efficient linear-programming-based algorithm for computing a factored value function for a factored MDP. We show how the use of factored value functions can form the basis for interdomain plan generalization, where we “learn” from plans constructed for some set of problems, and can provide good solutions to other unseen problems of the same type without any need for planning. We describe the application of this approach to the task of multi-agent planning in a strategic computer war game, and the application by Kok et al. for multi-agent coordination in their world-champion RoboSoccer team. Acknowledgements. Joint work with Carlos Guestrin, Ronald Parr, Chris Gearhart, Neal Kanodia, and Shobha Venkataraman.

Abstract: In multi-agent scenarios a planner must allocate agents to operators as well as instantiating operators to the entities appearing in the domain. For domains configured in this way the complexity and size of the search space increases dramatically as the number of possible actors increases. Using standard techniques this would result in an exponential explosion of the time required for the planner to find a plan. In this paper we present an extension to the Graphplan algorithm, which uses abstraction in order to avoid the complexity introduced by the increase in the number of possible actors in multi-agent planning. This planner does not commit to agent-action assignments during the search for a valid plan. Instead it generates a plan that is valid with respect to the number of actors available but is not committed to specific agent-action assignments. A post planning negotiation process can be used in order to obtain the commitment of agents to the performance of specific actions.

Abstract: : The fundamental problems addressed in this report pertained to the coordination of multi-agent software systems acting in dynamic, physical environments. The Multi-Agent Planning MAPLE project focused on the development and evaluation of new algorithms for integrating information and for coordinating the actions of large multi-agent teams. Three different aspects of multi-agent systems are in this report. First, this project developed a prototype of a secure decentralized database, which made it possible to disperse information among multiple agents in a secure and scalable way. Second, the project developed a multi-agent system for acquiring and integrating information in large dynamic environments. Third was a planning and multi-agent coordination technique for enabling large teams of autonomous agents to perform missions in highly dynamic and partially observable environments. The resulting algorithms were implemented and evaluated predominately using a simulator called "MapleSim", which simulated the country of Honduras after the devastation caused by Hurricane Mitch in 1998.