Batch A3C reinforcement learning method for agent exploring 3D maze

Abstract

The invention discloses a batch A3C reinforcement learning method for an agent to explore a 3D maze. In order to achieve the goal of relatively short training time and small memory loss, the inventionuses a batch-based reinforcement learning method to train a neural network. The neural network is divided into two parts. The first part mainly consists of several convolution layers and MLP, and thelow-dimensional representation of the original screen pixels is obtained. The second part is an LSTM (Long-Short-Term Memory) model. The input of the LSTM is the output of the MLP of the first part,and the cell output of the last time step of the LSTM is circumscribed with two MLPs, which are respectively used to predict the probability distribution of the action a in the current state and the prediction of the state value v in the current state. Combining the efficient reinforcement learning algorithm and depth learning method, the agent can explore 3D maze independently, and the agent cansuccessfully explore 3D maze environment with relatively short training time and small memory consumption.

Description

technical field [0001] The invention belongs to the field of reinforcement learning and deep learning, and mainly relates to a method for exploring a 3D maze by an agent based on deep reinforcement learning. Based on this scene, we can evaluate the training time of various deep reinforcement learning models, memory loss and the ability of the agent to explore the maze stability in the process. Background technique [0002] Reinforcement Learning (RL, Reinforcement Learning) is considered to be one of the core technologies for designing artificial intelligence systems. Reinforcement learning originated from the research of behavioral psychology, which largely imitates the learning mode of intelligent creatures, so that the agent (Agent) with reinforcement learning ability gradually learns the most effective interaction with the environment from its own experience. way, rather than telling the agent how to interact with the environment in advance. The agent's goal for each i...

AI Technical Summary

Problems solved by technology

However, in the continuous and complex environment of the reinforcement learning method, the state of the environment may be different every moment, and the bottleneck of the reinforcement learning theory will be revealed.

In the face of high-dimensional, dynamic programming problems with huge state sets, pure RL will appear powerless

Images