AlphaZero. It can be traced back to the TD-Gammon algorithm from

[rriiffaatt77]

Reinforcement learning (RL) is a feedback-based learning method that rewards and punishes correct and incorrect behavior that an algorithm makes the best decisions. The word "reinforcement" comes from psychology. "Reinforcement" in psychology is the establishment or encouragement of a pattern of behavior by providing a means of stimulation. This type of "reinforcement" is specifically divided into two types: Positive reinforcement refers to the provision of motivational stimuli to increase further positive responses after the expected behavior has been presented. Negative reinforcement corrects undesirable behavior by providing appropriate stimulation to reduce the likelihood of a negative (unwanted) response.



Imagine that when you play Super Mario for the first belgium email list time by yourself, you have to constantly explore the environment and important NPCs in the game to find a safe place to get gold coins! After n times of exploring rewards and punishments, you will become more and more skilled at Mario games, the correctness of operations will be greatly improved, and you will eventually become a master of the game. . Self-play Self-play is a synthetic method for learning algorithms such as unlimited power of AI to make up for the shortcomings of data usage.



efficiency. Taking AlphaZero as an example, in each game, the model uses Monte Carlo Tree Search (MCTS) to choose actions. MCTS combines the policy and value provided by the current neural networks to estimate the optimal action in each game state. The specific steps are as follows: ) Random initialization: The model starts from a completely randomly initialized state without any human prior knowledge. ) Self-play: The model plays against itself and generates a large amount of game data. Good results are used to update the model parameters. ) MCTS: In each game, AlphaZero will use MCTS to search for the best move.