Skip to content

Optim4RL is a Jax framework of learning to optimize for reinforcement learning.

License

Notifications You must be signed in to change notification settings

sail-sg/optim4rl

Repository files navigation

Optim4RL

This is the official implementation of Optim4RL, a learning to optimize framework for reinforcement learning, introduced in our RLC 2024 paper Learning to Optimize for Reinforcement Learning.

Table of Contents

Installation

  1. Install JAX 0.4.19: See Installing JAX for details. For example,
pip install --upgrade "jax[cuda12_pip]==0.4.19" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
  1. Install other packages: see requirements.txt.
pip install -r requirements.txt
  1. Install learned_optimization:
git clone https://github.com/google/learned_optimization.git
cd learned_optimization
pip install -e . && cd ..

Usage

Hyperparameter

All hyperparameters, including parameters for grid search, are stored in a configuration file in the directory configs. To run an experiment, a configuration index is first used to generate a configuration dict corresponding to this specific configuration index. Then we run an experiment defined by this configuration dict. All results, including log files, are saved in the directory logs. Please refer to the code for details.

For example, run the experiment with configuration file a2c_catch.json and configuration index 1:

python main.py --config_file ./configs/a2c_catch.json --config_idx 1

To do a grid search, we first calculate the number of total combinations in a configuration file (e.g. a2c_catch.json):

python utils/sweeper.py

The output will be:

The number of total combinations in a2c_catch.json: 2

Then we run through all configuration indexes from 1 to 2. The simplest way is using a bash script:

for index in {1..2}
do
  python main.py --config_file ./configs/a2c_catch.json --config_idx $index
done

Parallel is usually a better choice to schedule a large number of jobs:

parallel --eta --ungroup python main.py --config_file ./configs/a2c_catch.json --config_idx {1} ::: $(seq 1 2)

Any configuration index with the same remainder (divided by the number of total combinations) should have the same configuration dict (except the random seed if generate_random_seed is True). So for multiple runs, we just need to add the number of total combinations to the configuration index. For example, 5 runs for configuration index 1:

for index in 1 3 5 7 9
do
  python main.py --config_file ./configs/a2c_catch.json --config_idx $index
done

Or a simpler way:

parallel --eta --ungroup python main.py --config_file ./configs/a2c_catch.json --config_idx {1} ::: $(seq 1 2 10)

Please check run.sh for the details of all experiments.

Experiment

  • Benchmark classical optimizers: run a2c_*.json or ppo_*.json.
  • Collect agent gradients and parameter updates during training: run collect_*.json.
  • Meta-learn optimizers and test them:
    1. Train optimizers by running meta_*.json. The meta-parameters at different training stages will be saved in corresponding log directories. Note that for some experiments, more than 1 GPU/TPU (e.g., 4) is needed due to a large GPU/TPU memory requirement. For example, check meta_rl_catch.json.
    2. Use the paths of saved meta-parameters as the values for param_load_path in lopt_*.json.
    3. Run lopt_*.json to test learned optimizers with various meta-parameters. For example, check lopt_rl_catch.json.

Analysis

To analyze the experimental results, just run:

python analysis_*.py

Inside analysis_*.py, unfinished_index will print out the configuration indexes of unfinished jobs based on the existence of the result file. memory_info will print out the memory usage information and generate a histogram to show the distribution of memory usages in the directory logs/a2c_catch/0. Similarly, time_info will print out the time information and generate a histogram to show the time distribution in the directory logs/a2c_catch/0. Finally, analyze will generate csv files that store training and test results. Please check analysis_*.py for more details. More functions are available in utils/plotter.py.

Citation

If you find this work useful to your research, please cite our paper.

@article{lan2024learning,
  title={Learning to Optimize for Reinforcement Learning},
  author={Lan, Qingfeng and Mahmood, A. Rupam and YAN, Shuicheng and Xu, Zhongwen},
  journal={Reinforcement Learning Journal},
  volume={1},
  issue={1},
  year={2024}
}

License

Optim4RL is distributed under the terms of the Apache2 license.

Acknowledgement

We thank the following projects which provide great references:

Disclaimer

This is not an official Sea Limited or Garena Online Private Limited product.