Interactive Reinforcement Learning Strategy
Zhenjie Shi, Wenming Ma, Shuai Yin, Hailiang Zhang, Xiaofan Zhao · 2021
The birth of AlphaGo has set off a new wave of reinforcement learning technology. Reinforcement learning has become one of the most popular directions in the field of artificial intelligence. Its essence is the continuous integration and upgrading of various machine learning methods, and the agents continue to trial and error and obtain cumulative rewards. Q-learning is the most commonly used method in reinforcement learning, but it itself has many problems such as less early information, long learning time, low learning efficiency, and repeated trial and error. Therefore, Q-learning cannot be directly applied to the real environment. In response to this problem, the reinforcement learning discussed by the author is an interactive learning method that combines voice commands and Q-learning. This method uses part of the interaction between the agent and the human voice to find a larger target range in the early stage of learning. Then narrow the search range in turn, which can guide the agent to quickly achieve the learning effect and change the blindness of learning. Simulation experiments show that compared with the standard Q-learning algorithm, the proposed algorithm not only improves the convergence speed, shortens the learning time, but also reduces the number of collisions, enabling the agent to quickly find a better collision- free path.