I. Overview
[Reinforcement learning Reinforcement Learning, RL, enhancing learning] This term comes from behavioral psychology, biological represents more frequently for avoiding disadvantages and implement strategies to their advantage. [Including strengthening positive reinforcement positive reinforcement and negative reinforcement]] [negative reinforcement, positive reinforcement which makes biological tend to get more benefits, negative reinforcement such that the biological tendency to avoid damage. AI [Artificial Intelligence, AI] areas have many similar problems while avoiding disadvantages. For example, the famous AI program AlphaGo Go chess can be different according to different situations go. If it's good at, it will win; well, it will lose the next. It is based on the continuous improvement of their chess chess experience. This behavior and psychology of the situation is exactly the same. So, artificial intelligence borrowed the concept of behavioral psychology, the interaction with the environment while avoiding disadvantages of the learning process called reinforcement learning.
II. Reinforcement Learning and its key elements
1. In the field of artificial intelligence, reinforcement learning is a particular kind of machine learning problems. In a reinforcement learning system, decision-makers can observe the environment, and make actions based on observations. After the action, you can be rewarded. Reinforcement learning how to maximize rewards through interaction with the environment to learn.
2. reinforcement learning greatest feature is no correct answer in the learning process, but to learn through reward signal.
3. A reinforcement learning system has two key elements: incentives and policies.
"[Reward] rewards: the reward is to strengthen the learning objectives learning system. After the learner will receive environmental action sent the reward, and strengthen the goal of learning is to maximize the total reward value in the long time.
"Strategy [policy]: decision-makers will adopt different actions depending on the observation decision, which called the policy from the relationship between the observed action. Reinforcement learning objects learning is strategy. Reinforcement learning through improved strategies in order to maximize the total reward. Policy may be deterministic, it can be uncertainty.
4. The difference reinforcement learning and supervised learning
"For the supervised learning, the learner know what the correct answer to each action that can be learned by comparing gradually; for reinforcement learning, learners do not know the correct answer to each action can only be learned through reward signal. Reinforcement learning to maximize reward in a period of time, need to focus on longer-term performance. At the same time, he supervised learning hoping to use the results to learn to unknown data, the results can be generalized requirements, can be generalized; learning is enhanced results can be used in training environments. So, generally used to determine supervised learning, forecasting and other tasks, such as determining the content of the picture, forecasting stock prices; and reinforcement learning not to use such a task.
5. The difference reinforcement learning and unsupervised learning
"Unsupervised learning aimed at discovering hidden between data structures, and reinforcement learning has a clear numerical target, namely reward. Their different research purposes. So, generally used for unsupervised learning tasks such as clustering, and reinforcement learning do not apply to such a task.
III. Application of enhanced learning
1. Video Games
"Mainly refers to video game players need to operate in accordance with the contents of the screen of the games, including Pac-Man console games, PC game StarCraft, mobile games and other stimulating battlefield. Many games require the highest possible points, or to obtain victory in multi-party confrontation. At the same time, for these games, it is difficult to obtain in standard answer how each step should be operated. From this perspective, these games AI games require the use of reinforcement learning.
2. Board games
"Go, flipping black and white chess, backgammon and so on. AI can be achieved through a variety of board sports reinforcement learning. AI board has a clear goal - to improve the odds, but every step is often no absolute right answer. Notably the AlphaGo, AlphaZero etc.
3. autopilot