Sushant-ctrl/DQN-Algorithms

This Repository has the Implementations of variants of the Q Learning Algorithms

DQN BASED ALGORITHMS

Reinforcement Learning is all about finding the best strategy to do a particular task. These algorithms involve estimating the goodness of the position of the agent in an environment concerning its goal.

What is Q value?

Q-values are defined for states and actions. is an estimation of how good is it to take the action at the state.

Finding the Q values

Q values can be found using the bellman optimality equation.

But this repostiory isn't about using this method!!

Here instead of using the conventional method of sampling from the environment and updating the Q values, we use function approximators particularly Neural Networks, thus the name Deep Q Networks.

There are various variants to this algorithm I shall be implementing the following

• DQN with experience replay
• Double DQN
• Duelling DQN
• Noisy DQN
• Rainbow DQN
• DQN with HER

These algorithms are implemented on low state representation of state space.

Repository Structure

                |Readme.md
                |---DQN with experience replay
                |---Double DQN
                |---Duelling DQN
                |---Noisy DQN
                |---Rainbow DQN
                |---DQN with HER

Each subfolder is structured as

                |Readme.md
                |---Main.py
                |---Solver.py
                |---UTILS.py
                |---Running Trained Model.py
                |---Trained Model.pt