基于改进深度强化学习的三维环境路径规划

提出一种改进深度强化学习算法(NDQN),解决传统Q-learning算法处理复杂地形中移动机器人路径规划时面临的维数灾难。提出一种将深度学习融于Q-learning框架中,以网络输出代替Q值表的深度强化学习方法。针对深度Q网络存在严重的过估计问题,利用更正函数对深度Q网络中的评价函数进行改进。将改进深度强化学习算法与DQN算法在同样的三维环境下进行仿真实验,从最优路径长度、损失函数值、得到稳定的奖励值、收敛速度等方面进行对比,改进深度强化学习算法比DQN算法得到很大的改善,说明改进的算法比DQN算法寻得了更优的策略。

3D ENVIRONMENT PATH PLANNING BASED ON IMPROVED DEEP REINFORCEMENT LEARNING

This paper proposes an improved deep reinforcement learning algorithm(NDQN)to solve the dimensionality disaster faced by the traditional Q-learning algorithm in dealing with mobile robot path planning in complex terrain.A deep reinforcement learning method that incorporates deep learning into the Q-learning framework and replaces the Q-value table with network output was proposed;due to the serious overestimation problem for the depth Q network,a correction function was used to improve the evaluation function in the deep Q network.The improved deep reinforcement learning algorithm and DQN algorithm were tested in the same 3 D environment.The optimal path length,the loss function value,the stable reward value and the convergence speed were compared.The improved deep reinforcement learning algorithm is much better than the DQN algorithm.It shows that the improved algorithm finds a better strategy than the DQN algorithm.