WilsonWangTHU / mbbl-slbo

Stochastic Lower Bound Optimization

Algorithmic Framework for Model-based Deep Reinforcement Learning with Theoretical Guarantees. 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.

This is the implementation built on top of the official repo by the authors.

Installation

Requires mujoco131.

Install required packages using:

pip install -r requirements.txt

Then please go to MBBL to install the mbbl package for the environments.

Run

To do the benchmarking, please refer to ./example_script/run_experiments.sh First copy the script under the root directory, i. e. cp ./example_script/run_experiments.sh ./ Then run the benchmarking for example as

bash run_experiments.sh gym_cheetah

Before running, please make sure that rllab and baselines are available

python main.py -c configs/algos/slbo.yml configs/envs/half_cheetah.yml -s log_dir=/tmp

Environments are modified to use the gym environments for the benchmarking project. Supported environments are:

Environment         |   Max timesteps
---------------------------------------
reacher             |   50
half_cheetah        |   1000
walker              |   1000
hopper              |   1000
swimmer             |   1000
ant                 |   1000
                    |
pendulum            |   200
inverted_pendulum   |   100
acrobot             |   200
cartpole            |   200
mountain            |   200

If you want to change hyper-parameters, you can either modify a corresponding yml file or change it temporarily by appending model.hidden_sizes='[1000,1000]' in the command line.

License

See LICENSE for additional details.