Extending an Estimated-Regression Planner for Multi-Agent Planning

Contents

We examine the issues that arise in extending an estimated-regression planner to find plans for multiagent teams, cooperating agents that take orders but do no planning themselves. An estimated-regression system is a classical planner that searches situation space, using as a heuristic numbers derived from a backward search through a simplified space, summarized in the regression-match graph. Extending the planner to work with multiagent teams requires it to cope with autonomous processes, and objective functions that go beyond the traditional step count. Although regressing through process descriptions is no more difficult than regressing through standard action descriptions, figuring out how good an action recommended by the regression-match graph really is requires projecting the subtree suggested by the action. We are in the process of implementing the algorithm.