Comparisons of Different Dynamic Game Reinforcement Learning Optimal Strategies for A Non-holonomic Vehicle

In this paper, we compare different reinforcement learning dynamic (differential) game strategies for a nonholonomic vehicle in the linear-quadratic optimal control problem formulations. We consider conflict games Nash and Stackelberg and cooperative Pareto games. Interesting results are obtained that indicate when the conflict (Nash or Stackelberg) is better than cooperation (Pareto). We draw the boundaries when the controllers (players) prefer to play Nash strategies, when to play Stackelberg strategies, and when to cooperate and play Pareto strategies. The decisions are made on the optimal quadratic performance values for each player and for the parameter of cooperation (or no cooperation, when it is equal to zero).