Skip to main content
OverflowAI is here! AI power for your Stack Overflow for Teams knowledge community. Learn more

Reinforcement learning is a technique wherein an agent improves its performance via interaction with its environment. For this reason it is a commonly used machine learning technique in robotics.

Reinforcement learning (RL) is a machine learning (ML) technique wherein an agent improves its performance through interaction with its environment by attempting to minimize an objective or cost function referred to as the reward function.

RL is a class of algorithms separate from both supervised learning and unsupervised learning. It differs from supervised learning methods in that RL agents are not provided with the correct answer at the time of training. Instead they receive a "reward" with each action. This reward is often referred to as a penalty when its value is negative. Furthermore the reward function also defines how RL differs from unsupervised learning methods. Specifically unsupervised learning methods receive no information what-so-ever regarding what is correct during training.

Common formulations of the reward function provide the agent with a positive or zero reward as long as it performs within the acceptable bounds and a negative reward if it performs incorrect or sub-optimal actions.

Frequently an RL agent cannot directly determine the reward function and instead it develops a policy that defines which action to take given a particular state of the world. This policy is developed through repeated training and following the policy maximizes the reward.

One problem with always following the policy is that nothing new can be learned in doing so. Since there is no guarantee that the current policy is optimal it is necessary for the agent to occasionally select actions other than what is prescribed by the policy. This is known as exploration and permits the agent to find superior solutions. However it is done at the risk of receiving penalties or lower overall rewards. One common technique for selecting actions off-policy is epsilon-greedy selection. This method has the agent randomly select an action some percentage of the time and select the policy action the remaining percentage of the time.

The text Reinforcement Learning: An Introduction by Sutton and Barto details the myriad facets of reinforcement learning.