This project implements a reinforcement learning approach based on deep Q-learning to train a neural policy for navigating and solving reward network tasks.
The algorithm simulates how agents explore networks by learning to maximize rewards while navigating through nodes. It mimics intuitive or heuristic decision-making by only considering immediate options without explicit planning.
- Trains a model-free neural policy capable of solving the reward network tasks.
- Evaluates performance against heuristic benchmarks, such as random agents and myopic strategies.
- Uses unique network environments for training and testing to ensure robustness.
See the Main README for the setup of the project.
The training of the DQN agent requires a dataset of networks. To generate the networks, run the following command:
docker compose run all python common/generate/generation.py -i config/networks_train.yml -o data/networks_train.jsonTo train the DQN agent on the generated networks (training dataset), run the following command:
docker compose run all python algorithm/dqn/dqn_agent.py --config algorithm/params/seed_0.ymlTo apply the trained DQN agent on the generated networks (experiment dataset), run the following command:
docker compose run all python algorithm/dqn/dqn_exp.py --config algorithm/params/seed_0.yml