This application implements a simple framework for solving the problems available in the OpenAI gym framework.
The genetic algorithm is used to optimize the weight and bias matrices of the agent's neural network.
The application requires Python 3.
To setup the project and install the necessary dependencies, simply run:
make install
To run the application in training mode, simply use the following commmand:
make train
or, if you want to be more specific in the training process, try the following:
python3 src/main.py python src/main.py --mode=train --task=cartpole --population_size=200 --mutation_rate=0.07 --max_generations=20 --np=4 --crossover='uniform'
All these parameters are at your disposal. Note, that you can parallelize the training process by specifying the --np option.
To visualize one of the enclosed trained agent models, simply use the following command:
make replay path=[path to the trained agent pickle]
make replay path=./agents/walker.pkl
make replay path=./agents/cartpole.pkl