Converting MA-PDDL to extensive-form games

Abstract

This paper presents algorithms for converting multi-agent planning (MAP) problems described in Multi-Agent Planning Domain Definition Language (MA-PDDL) to extensive-form games in order to analyse and solve them with game-theoretic tools in general. MA-PDDL is an attempt to standardize the description of MAP problems similarly to PDDL in the single-agent setting. In this paper MA-PDDL is extended with partial-observability and probabilistic-effects to model more realistic domains. The conversion is fruitful in both ways: 1) extensive-form games can be solved via game-theoretic solution concepts (e.g. Nash-equilibrium) providing solutions to corresponding MAP problems in general, and 2) MA-PDDL problems can be solved via MAP methods providing solutions to corresponding games. Both cooperative and non-cooperative solutions can be achieved.