REINFORCE is a policy based RL algorithm. It collects couple of episodes and then generate a loss function to be back-propagated along the network. I mostly use it for continous/discrete state space but only discrete action space. Algorithm :

• For an episode, we generate a categorical distribution of probabilites and then collect the log probabilities of each action taken.
• After an episode is over, we mulitply the log probabilities with the associated reward for that action.
• The main aim is to give more preference to those actions which generates more reward.
• The loss function is calculated by taking the sum of all the product of log probabilites with total reward.

Projects solved:

• 1 - CartPole - For info on the environment - https://github.com/openai/gym/wiki/CartPole-v0
• 2 - LunarLander - For info on the environment - https://github.com/openai/gym/wiki/Leaderboard#lunarlander-v2
• 3 - Acrobot - For info on the environment - https://gym.openai.com/envs/Acrobot-v1/