Super Mario Bros using Double Deep Q Learning

This repository contains the code for an agent that can play Super Mario Bros Game using Double Deep Q Learning Algorithm.

Result

Double Deep Q Learning Algorithm

Double Deep Q Learning reduces the overestimation occured in Deep Q Learning by decomposing the max operation in the target into action selection and action evaluation.

This is done by using two independent agents, Online Agent and Target Agent. The Online Agent is trained for few episodes then the parameters of Online Agent is used to update the parameters of Target Agent this stabilizes the training process of the agents.

The environment data or state is first preprocessed and made ready for the agents to train them on it. Now following the exploitation and exploration trade off, a random number between 0 to 1 is generated if the number is below epsilon value then a random action is performed for the given state else the state is passed to the Online Agent to predict which action has the highest Q value, the action with the highest Q value is performed. At first the value of epsilon is 1 which means the agents are only going to explore and not exploit at lot and after each episode the value of episolon is decreased by a weight of exploration decay rate, so that the agents can slowly start exploiting in the environment.

The next_state, reward and a boolen done (indicating if the game is ended or not) is returned by the Super Mario Bros Game Environment, then the state, next_state, reward and done is added to a memory which is implemented using a Deque.

From the memory randomly batch_size number of data is picked, each data consist of state, next_state, reward and done. which is passed to the Online Agent and Target Agent for generating their corresponding Q Values. A loss is generated using Smooth L1 Loss function over the Q Values of Online Agent and Target Agent. Then this loss is used to backpropagate over the Online Agent.

This entire process from preprocessing the information data or state of Super Mario Bros Game Environment to backprogration of Online Agent is called an episode. Multiple episodes are performed in order make the agents learn an idle Policy Function. Here I have trained the models for 40000 episodes and on a exploration decay rate of 0.99999975.

After every interval of 10000 episodes the weights of the Target Model is synced with the weights of the Online Agent.

Preprocess

The agent perceives the Super Mario Bros Game Environment data or state frame by frame as an RGB image of size [3, 240, 256]. Here there is more information than the agent requires, for instance Mario's actions do not depend on the color of the pipes or the sky.

So, we have to preprocess the environment data to remove the unwanted information this can be done by implementing Wrappers over the Super Mario Bros Game Environment. Here the environment data, an RGB image of size [3, 240, 256] is passed to the GrayScaleObservation wrapper that converts it into a grayscale image of size [1, 240, 256], this is then passed to ResizeObservation wrapper which downsamples the graysacle image to size [1, 84, 84]. The downsampled image is then passed to SkipFrame wrapper which skips n consecutive frames because they don't vary much. The nth frame aggregates rewards accumulated over each skipped frame. Then a FrameStack wrapper is applied to squash consecutive frames of the environment into a single observation point to feed to our learning model.

Training

To Train the Agent use the following command :

python3 train.py --world world_number --stage stage_number --checkpoints /directory/to/save/the/trained
/agents/ --episodes episodes

Following are the details for the commnad line arguments passed in the training command:

1. --world : this argument is used to specify on which world number the agents will be trained, here world_number represents the world number in which the agents will be trained. There are total 8 world in Mario Bros Game you can specify any number between 1 to 8.

2. --stage : this argument is used to specify on which stage number of the specified world the agents will be trained, here stage_number represents the stage number of the specified world where the agents will be trained. There are total 4 stages in every world of Mario Bros Game you can specify any number between 1 to 4.

3. --checkpoints : this argument is used to specify the directory path where the trained models will be saved. Here /directory/to/save/the/trained represents the directory path where the trained agents will be saved.

4. --episodes : this argument is used to specify the number of episodes till which the agents will be trained. Here episodes represents the number of episodes till which the agents will be trained.

Testing

To Test the Agent use the following command :

python3 test.py --world world_number --stage stage_number --checkpoints /path/to/the/checkpoints/file/of/the/trained/agents/ --episodes episodes

Following are the details for the commnad line arguments passed in the testing command:

1. --world : this argument is used to specify on which world number the trained agent will play, here world_number represents the world number in which the traind agent will play. There are total 8 world in Mario Bros Game you can specify any number between 1 to 8.

2. --stage : this argument is used to specify on which stage number of the specified world the trained agent will play, here stage_number represents the stage number of the specified world where the trained agent will play. There are total 4 stages in every world of Mario Bros Game you can specify any number between 1 to 4.

3. --checkpoints : this argument is used to specify the path where the checkpoints file of the trained agent is saved. Here /path/to/the/checkpoints/file/of/the/trained/agents/ represents the path where the checkpoints file of the trained agent is saved.

4. --episodes : this argument is used to specify the number of times the trained agent should play then game. Here episodes represents the number times the trained model should play the game.

Requirements

To install the required dependencies for this project use the following command :

pip3 install -r requirements.txt

This command will install the following python packages :

gym==0.25.2
gym_super_mario_bros==7.4.0
matplotlib==3.7.1
nes_py==8.2.1
numpy==1.23.5
scikit_image==0.19.3
torch==2.0.1+cu118