Different types of reinforcement learnings will be covered here.

I begin with model based learning and show solutions for FrozenLake using OpenAI gym using value iteration and policy iteration.

Conceptually in value iteration we do the following:

• compute the optimal value function first for each state-action pair
• extract the optimal policy from the optimal value function

While using the policy iteration method we do the following:

• start with a random policy
• compute the value function
• extract a new policy using the value function from the previous step
• compare the old and the new policy and stop if the difference between them is below a threshold, otherwise continue with another policy (and compute the value function again)

For model free based problems, solutions for Monte Carlo prediction methods, temporal differencing (e.g.: Q Learning), multi armed banidts and deep q network are implemented. The different models are tested in jupyter notebooks where results can be observed.