Reinforcement Learning Model Based On Optical Neural Network And Its Application

In recent years,optical neural network has attracted extensive attention in the field of artificial intelligence due to its advantages of high speed,low power consumption and large bandwidth.However,most of the current optical neural networks are used in supervised learning tasks,and supervised learning needs a lot of prior data to support,which also makes the application scenarios of supervised learning more limited.With the rapid development of the field of artificial intelligence,people also put forward higher requirements for the learning ability of the agent.It needs the agent to learn more complex tasks through the constant interaction with the environment,which is reinforcement learning.In this paper,optical neural network is applied to reinforcement learning task,which not only expands the application field of optical neural network,but also provides a new idea for the implementation of reinforcement learning.1.This paper proposes an optical neural network reinforcement learning model(ORL)based on optical neural network,and applies the ORL model to the reinforcement learning environment of discrete data space.In this paper,we use grid world,which is a common environment model in reinforcement learning,and build two-dimensional and three-dimensional situations independently.In the two-dimensional grid world,the agent based on ORL model can stably find the shortest path to the end after 200 games;in the three-dimensional grid world,the agent based on ORL model can walk to the end of the maze after 300 games from 1000 steps in each game to 20 steps in each game.2.The ORL model is applied to the continuous data space reinforcement learning environment.The continuous state environment used in this paper is the cart pole game in the reinforcement learning environment toolbox open AI gym.The simulation results show that the agent based on ORL model can stably reach the maximum step size of each game after 450 games,that is,the car can stably support the slider without tipping.Through the performance comparison between ORL model and DQN,we can conclude that the agent based on ORL model has the same learning ability and environment adaptability as the electronic reinforcement learning architecture,which also proves that the application of optical neural network in reinforcement learning has a very attractive development prospect.