Event

A Dynamic Game Learning Framework for Collaborative Iterative Tasks

1 reports · 1 sources

Mark Pustilnik and Francesco Borrelli propose a general learning framework for nonlinear, non-convex dynamic games, addressing multi-agent collaborative iterative tasks with shared constraints by constructing a sampled joint safe set and each participant's cost-to-go function from executed feasible trajectories. Under this framework, agents jointly select at each step a sampled terminal state and a generalized Nash equilibrium solution reaching it; a new solution is adopted only if it satisfies a collaborative acceptance rule agreed upon by the participants. The authors prove that the framework guarantees recursive feasibility and provide conditions under which each participant's cost does not increase across iterations.