General Field and Reinforcement Learning RL

Author: Zen and the Art of Computer Programming

1 Introduction

(1 Introduction

More than two years have passed in 2019. In the field of artificial intelligence, there has been very vigorous development in both research and application levels. In recent years, reinforcement learning (Reinforcement Learning, RL) has become one of the hottest directions of artificial intelligence, and its own unique characteristics and applications have increasingly attracted widespread attention. Due to its powerful learning ability, it can effectively solve complex decision-making problems, control robots, and enhance game AI.

Reinforcement learning is to let the agent (Agent) continuously adjust the policy (Policy) to achieve the optimal state and finally achieve the goal by interacting with the environment. Its core task is how to let the agent find the best strategy in the process of continuous exploration and learning. It has good applications in many fields such as robot control, automatic driving, reinforcement learning, game AI, etc.

Reinforcement learning is mainly divided into four stages:

1. Agent search strategy (Exploration): By exploring and discovering new strategies, the agent can better adapt to the environment. Typical algorithms include stochastic strategies, model strategies, and methods based on policy gradients.

2. Agent learning strategy (Learning): Learn new strategies from historical data, so that the agent can better execute the strategy in the current environment. Typical algorithms include Q-learning, SARSA, etc.

3. Time to State Representation: Convert time to state representation, which is convenient for training and RL algorithms. Typical tools are vector products, matrix representations, etc.

4. Reward Function Design: The reward function refers to the reward value obtained by giving the agent to perform a specific action, and is used to measure the performance of the agent. Designing an appropriate reward function according to the actual situation can improve the learning efficiency of the agent. Typical reward functions include action-based and value-based rewards, etc.

Reinforcement learning belongs to a comprehensive research field, involving computer science, economics, mathematics, engineering, statistics and other disciplines. This article will introduce the following aspectsÿ