/ artificial intelligence

Reinforcement Learning

Reinforcement learning is type of machine learning that is focused on maximisation of reward received by interactions with environment. This idea isn't new actually research on this topic where started in early 80's. First versions where based on tables of probability. The idea of learning through interactions with environment in probably closest to human learning. Reinforcement learning (RL) is inspired by behaviourist psychology. The main goal of this kind of learning is to maximize numerical reward returned by critic. The learner is taking actions to determine which action or combination of actions yelds the most reward. RL is testing by trial-and-error and can handle delayed rewards.

In Machine Learning there are 3 types of learning:

  • supervised learning
  • unsupervised learning
  • reinforcement learning

Reinforcement learning is something in the middle between supervised and unsupervised because it don't have any knowledge what action should it take in concrete state but it has information if actions where better or worse than previous.

Main elements of reinforcement learning

  • policy
  • reward signal
  • value function
  • model*
policy

Policy is defining agents way of interacting with environment. We can say it is a mapping of environment states to actions. Sometimes we just can't make some actions.

reward signal

This is immediet value that agent receives for every action it make. In some cases there is no immediet reward. For example in chess we can't evaluate any single action, only final result of match is relevant.

value function

Value function is for determining which action made in specific state can give us the highest reward in the end. This is the total ammount of revard that agent can expect to accumulate over the future.

model

This is a model of the environment. This element is optional and can be ommited. Model of the environment gives agent ability to predict next state or/and reward signal after taking an action.

Resources

[1] Richard S. Sutton and Andrew G. Barto. 2017. Reinforcement Learning: An Introduction (2st ed.)