Amirhossein Afkhami Ardekani Combining deep learning and game theory for path planning in autonomous racing cars

Amirhossein Afkhami Ardekani, Amirhosein Chahe, Mohammad Reza Hairi Yazdi

Abstract: In this paper, a novel algorithm based on Nash equilibrium and memory neural networks has been suggested for the path selection of autonomous vehicles in highly dynamic and complex environments. The proposed algorithm has been investigated in a racing game containing two vehicles: ego and opponent. However, the proposed method can be easily generalized to a race game with more than two cars. The suggested method, called GT-LSTM, is comprised of three primary parts. First, memory neural networks for learning and predicting the opponent’s vehicle behavior. Second, game theory and payoff matrices which agents use to select the action that leads to the maximum payoff, and finally, PID controllers in order to make lane changing and path following smoother. The motivation behind combining game theory and memory neural networks was to facilitate and make the decision-making process for the ego vehicles more accurate by learning the opponent vehicle’s behavior through memory neural networks. In the designed scenario for investigating the effectiveness of the proposed method in the CARLA simulator, the opponent vehicle performs extreme maneuvers to block the ego’s route. On the other hand, the ego vehicle has the privilege of using the GT-LSTM method for overtaking the opponent and winning the race. For that, it uses the predicted action of the opponent vehicle by means of memory neural networks to update its own payoff matrices and opt for the best action, knowing the next move of its rival. For better performance comparison of the proposed algorithm, two other planning methods, including game theory without having knowledge about the opponent vehicle and also game theory and payoff matrices along with conventional neural networks, have been implemented. With respect to the yielding results from the simulation, the proposed method is superior to the two other methods, having a success rate of 55 percent, compared to 15 and 32 percent, respectively, for the game theory without having knowledge about the opponent vehicle and also game theory and payoff matrices next to conventional neural networks. Moreover, it has been shown that the obtained response from the suggested algorithm matches with Nash equilibrium in 90.2 percent of the situation during the simulation experiments.