WilsonWangTHU / mbbl-GPS

Guided Policy Search - GPS (MDGPS)

This repo is based on the original GPS. We modified the repo to perform benchmarking as part of the Model Based Reinforcement Learning Benchmarking Library (MBBL). Please refer to the project page for more information.

Installation

For full documentation, see rll.berkeley.edu/gps. First install this library as instructed in rll.berkeley.edu/gps. Then please go to MBBL to install the mbbl package for the environments.

Run the code

To disable the rendering and run the experiments on headless servers. first run

xvfb-run -s "-screen 0 1400x900x24" bash

Perform benchmarking

Please refer to exp_script/gym_search_2.sh.

An example to run HalfCheetah looks like this:

env_name=gym_cheetah
# python python/gps/gps_main.py gym_cheetah_mdgps_example 2>&1 | tee output.log
for batch_size in 5000; do
    for rand_seed in 1234 2345 2314 1234 1235; do
        # generate the config files
        exp_name=example_${env_name}_batch_${batch_size}_seed_${rand_seed}
        cp -r experiments/gym_mdgps_example experiments/${exp_name}_mdgps

        # modify the config files
        sed -i "s/ENV_NAME/${env_name}/g" experiments/${exp_name}_mdgps/hyperparams.py
        sed -i "s/RAND_SEED/${rand_seed}/g" experiments/${exp_name}_mdgps/hyperparams.py
        sed -i "s/TIMESTEPS_PER_BATCH/${batch_size}/g" experiments/${exp_name}_mdgps/hyperparams.py

        # run the experiments
        python python/gps/gps_main.py ${exp_name}_mdgps 2>&1 | tee ./log/${exp_name}_mdgps.log

    done
done

The configuration can be changed by modifying the template config file experiments/gym_mdgps_example.