Selective Determinism for Autonomous Navigation in Multi-Agent Systems

Abstract

Standard approaches to multi-agent navigation problems formulate them as searches for policies that are optimal mappings from belief states to actions. However, computing such policies is almost always intractable, both in theory and in practice, due in part to the combinatorial effects of reasoning about uncertain interactions into the future. This dissertation proposes a framework to address that intractability by identifying when and how interaction effects can be factored out of the problem while maintaining collision guarantees and goal-directed motion. At a low level, stochastic optimal control theory is leveraged to formulate a constrained interference minimization principle within which multi-objective control problems can be formulated and solved to a defined level of confidence. At a high level, it is shown that, under certain conditions, complex multi-agent decision process problems can be factored into independent sub-problems, which removes coordination effects and greatly reduces overall complexity. These two results are unified into a single problem solving strategy called the Selective Determinism (SD) framework, which enables robust and efficient solutions to multi-agent navigation problems.