Selective Negotiations for Scaling Stochastic Dynamic Games
Kamran Vakil, Alyssa N. Pierson · 2023
This paper presents asocial agents for selective negotiations in stochastic dynamic games. Game-theoretic frameworks in the belief-space show promise in modeling complex interactions in scenarios such as surveillance, herding, and racing. Stochastic dynamic games can be solved as a continuous POMDP to find a local Nash Equilibrium solution of all agents using a game-theoretic belief-space variant of iLQG. However, the scalability of this method suffers due to the large dimensionality of beliefs which the iLQG must propagate, and fails to consider environmental features such as dynamic obstacles. We introduce asocial agents models, which follow fixed input trajectories in the planning horizon. Ego agents can selectively toggle which agents it considers asocial, thereby reducing computational overhead. Simulations demonstrate that our approach to selective negotiations can help scale stochastic dynamic games with faster computation times and minimal performance decline.