This repository contains the source code for the paper titled More Efficient Randomized Exploration for Reinforcement Learning via Approximate Sampling.

• Python: >= 3.8
• Tianshou: ==0.4.10
• Envpool: ==0.6.6
• Additional dependencies can be found in requirements.txt.

Hyperparameters and grid search parameters are organized within a configuration file located in the configs folder. To initiate an experiment, select a configuration index to generate a corresponding dictionary. This dictionary defines the specific experiment setup. All outputs, including logs, are stored within the logs folder. For detailed instructions, refer to the provided source code.

To launch an experiment using the configuration file atari8_fg_aULMC.json with the index 1, execute:

python main.py --config_file ./configs/atari8_fg_aULMC.json --config_idx 1

To identify the total number of parameter combinations for a given configuration (for instance, atari8_fg_aULMC.json), run:

python utils/sweeper.py

This command outputs the total combinations:

Number of total combinations in atari8_fg_aULMC.json: 1728

To systematically explore each combination (indices 1 to 144), you could utilize a bash script:

for index in {1..144}
do
  python main.py --config_file ./configs/atari8_fg_aULMC.json --config_idx $index
done

For handling a large batch of experiments, GNU Parallel is recommended for job scheduling:

parallel --eta --ungroup python main.py --config_file ./configs/atari8_fg_aULMC.json --config_idx {1} ::: $(seq 1 1728)

If conducting multiple runs for the same configuration index, increment the index by the total number of combinations. For instance, to perform 5 runs for index 1:

for index in 1 1729 3457 5185 6913
do
  python main.py --config_file ./configs/atari8_fg_aULMC.json --config_idx $index
done

Alternatively, for simplicity:

parallel --eta --ungroup python main.py --config_file ./configs/atari8_fg_aULMC.json --config_idx {1} ::: $(seq 1 1728 8640)

To analyze experiment outcomes, simply execute:

python analysis.py

This script identifies unfinished experiments by checking for missing result files, reports memory usage, and produces a histogram of memory utilization for the logs/atari8_fg_aULMC/0 directory. It also generates CSV files summarizing the training and testing outcomes. For comprehensive details, see analysis.py. Additional analysis tools are available in utils/plotter.py.