Argumentation-based methods for multi-perspective cooperative planning

Abstract

Through cooperation, agents can transcend their individual capabilities and achieve goals that would be unattainable otherwise. Existing multiagent planning work considers each agent’s action capabilities, but does not account for distributed knowledge and the incompatible views agents may have of the planning domain. These divergent views can be a result of faulty sensors, local and incomplete knowledge, and outdated information, or simply because each agent has conducted different inferences and their beliefs are not aligned. This thesis is concerned with Multi-Perspective Cooperative Planning (MPCP), the problem of synthesising a plan for multiple agents which share a goal but hold different views about the state of the environment and the specification of the actions they can perform to affect it. Reaching agreement on a mutually acceptable plan is important, since cautious autonomous agents will not subscribe to plans that they individually believe to be inappropriate or even potentially hazardous. We specify the MPCP problem by adapting standard set-theoretic planning notation. Based on argumentation theory we define a new notion of plan acceptability, and introduce a novel formalism that combines defeasible logic programming and situation calculus that enables the succinct axiomatisation of contradictory planning theories and allows deductive argumentation-based inference. Our work bridges research in argumentation, reasoning about action and classical planning. We present practical methods for reasoning and planning with MPCP problems that exploit the inherent structure of planning domains and efficient planning heuristics. Finally, in order to allow distribution of tasks, we introduce a family of argumentation-based dialogue protocols that enable the agents to reach agreement on plans in a decentralised manner. Based on the concrete foundation of deductive argumentation we analytically investigate important properties of our methods illustrating the correctness of the proposed planning mechanisms. We also empirically evaluate the efficiency of our algorithms in benchmark planning domains. Our results illustrate that our methods can synthesise acceptable plans within reasonable time in large-scale domains, while maintaining a level of expressiveness comparable to that of modern automated planning.