**Q-Learning is a model-free, off-policy reinforcement learning algorithm that finds the best course of action given the current state of an agent.**Based on where the agent is in the environment, it will decide what action to take next. “Q” refers to the function that the algorithm calculates — the expected reward for an action taken in a given state.

The goal of Q-learning is to find the best course of action given the current state. To do this, it may develop its own rules or operate outside of a prescribed policy.**This means that no policy is actually needed, hence the term "off-policy".**For any finite Markov decision process, Q-learning finds an optimal policy that maximizes the expected value of the total reward at any and all consecutive steps, starting from the current state. Q-learning can identify the best action selection policy for any given finite Markov decision process, given infinite exploration time and a partially randomized policy.

An example of Q-learning is an ad recommendation system. In a normal ad recommendation system, the viewer gets ads based on the viewer’s previous purchases or websites that the viewer may have visited. If the viewer has purchased a TV, the viewer gets recommendations of different brands of TVs.

Important terms in Q-Learning

1. States: States S (States) represent the current position of the agent in the environment.
2. Action: An action is a step taken by an agent when it is in a specific state.
3. Reward: For each action, the agent receives a positive or negative reward.
4. Plot: When the agent ends up in a terminal state and cannot take new actions.
5. Q value: It is used to determine how good an action A is when it is executed in a specific state S. It is expressed as Q (A, S).
6. Temporal Difference: A formula for finding the Q value by using the current state and action and the values of the previous state and action.