A Reinforcement Learning Path Planning Method Using Artificial Potential Field

Abstract

The invention discloses a reinforcement learning path planning method by introducing an artificial potential field. Initialize the algorithm parameters; S3, use the dynamic factor adjustment strategy to select the action; S4, execute the action, update the Q value; S5, repeat the third and fourth steps until a certain number of steps or a certain convergence condition is reached; S6, each step The action with the largest Q value is selected to obtain the optimal path; S7, the optimal path is sent to the controller of the mobile robot, and the mobile robot is controlled to walk according to the optimal path. Compared with the traditional algorithm, the improved Q-learning algorithm in the present invention shortens the path planning time by 85.1%, reduces the number of iterations before convergence by 74.7%, and at the same time improves the stability of the convergence result of the algorithm.

Description

technical field [0001] The invention relates to the technical field of robot path planning, in particular to a reinforcement learning path planning method by introducing an artificial potential field. Background technique [0002] With the development of science and technology, more and more mobile robots have entered people's daily life. The path planning problem of mobile robots has also received more and more attention. The path planning technology can help the robot avoid obstacles to plan an optimal movement route from the starting point to the target point under the condition of referring to a certain index. According to the known degree of environmental knowledge in the path planning process, path planning can be divided into global path planning and local path planning. Among them, the widely used global path planning algorithms include A* algorithm, dijkstra algorithm, visual graph method, free space method, etc.; local path planning algorithms include artificial ...

AI Technical Summary

Problems solved by technology

The traditional Q-learning algorithm has the following problems: (1) During the initialization process, all Q values ​​are set to 0 or random values, which makes the agent only search blindly at the initial stage, resulting in There are too many invalid iterations in the stage; (2) ε-greedy strategy is adopted in the action selection, too large ε value will make the agent explore more environments and it is not easy to converge, and too small ε value will cause the agent to be sensitive to the environment. Finding a suboptimal solution due to insufficient exploration, it is difficult to balance the relationship between exploration and utilization

Images