Deep Reinforcement Learning Path Planning Algorithm Based on SARSA

In order to solve the problem that traditional path planning algorithms cannot adapt to complex environment, a deep reinforcement learning path planning algorithm based on SARSA is proposed. On the basis of deep SARSA networks (D-SARSA) algorithm of deep reinforcement learning, a goal-guided reward function is proposed to improve sparse reward problem. In order to speed up the convergence of the algorithm, the concept of important experience replay is proposed in combination with the idea of prioritized experience replay, and on this basis, deep SARSA networks with important experience replay (D-SARSA-IER) algorithm is proposed. In order to prove the effectiveness of the proposed algorithm, a helicopter raid radar position model is established, and a control experiment on path planning is conducted using deep Q-network (DQN), D-SARSA and D-SARSA-IER algorithms. During the experiment, the convergence speed of D-SARSA-IER algorithm is increased by 17.6% and 22.2% respectively compare with DQN and D-SARSA algorithm. In helicopter raid radar position task, the success rate of D-SARSA-IER algorithm is 9% higher than that of DQN algorithm, and the path planned by D-SARSA-IER algorithm is more stable.