This repo contains the PyTorch implementation of the PCPG algorithm, described in the following paper: PCPG: Policy Cover Directed Exploration for Provable Policy Gradient Learning, by Alekh Agarwal, Mikael Henaff, Sham Kakade and Wen Sun.
The necessary dependencies are in the environment.yml file. You can run
conda env create -f environment.yml
./setup.sh
To run PCPG on the Bidirectional Diabolical Combination Lock environment, execute:
python run.py -alg ppo-pcpg -env diabcombolockhallway -horizon 6 -lr 0.0005
The scripts submit_combolock_pcpg.sh, submit_combolock_ppo_rnd.sh and submit_combolock_ppo.sh will run the PCPG, PPO+RND and vanilla PPO algorithms on the combination locks for different horizon lengths. PCPG solves the task for all horizon lengths at least 95% of the time, while PPO+RND only succeeds roughly 50% of the time due to not adequately exploring both locks. Vanilla PPO fails due to the antishaped rewards.
For the Diabolical Combination Lock environment, the learned policies in the policy cover can be seen in the .txt files written to the output folder. For example, for horizon length 5, the policies and mixture weights are given by:
policy mixture weights: [0.36501506 0.40568325 0.22930169]
state visitations for policy 0:
lock1
[[0.048 0.004 0. 0. 0. ]
[0.033 0.004 0. 0. 0. ]
[0. 0.073 0.08 0.081 0.081]]
lock2
[[0.051 0.004 0. 0. 0. ]
[0.034 0.005 0. 0. 0. ]
[0. 0.077 0.085 0.086 0.086]]
state visitations for policy 1:
lock1
[[0.069 0.076 0.061 0.064 0.06 ]
[0.056 0.049 0.064 0.061 0.065]
[0. 0. 0. 0. 0. ]]
lock2
[[0.022 0.019 0.02 0.02 0.02 ]
[0.019 0.022 0.021 0.018 0.018]
[0. 0. 0. 0.003 0.004]]
state visitations for policy 2:
lock1
[[0.003 0.003 0.004 0.004 0.003]
[0.004 0.003 0.003 0.003 0.004]
[0. 0. 0. 0. 0. ]]
lock2
[[0.077 0.093 0.085 0.075 0.075]
[0.083 0.067 0.075 0.085 0.066]
[0. 0. 0. 0. 0.019]]
state visitations for weighted policy mixture:
lock1
[[0.046 0.033 0.026 0.027 0.025]
[0.036 0.022 0.027 0.025 0.027]
[0. 0.027 0.029 0.03 0.03 ]]
lock2
[[0.045 0.031 0.028 0.025 0.025]
[0.039 0.026 0.026 0.027 0.022]
[0. 0.028 0.031 0.033 0.037]]
The scripts submit_mountaincar_pcpg.sh and submit_mountaincar_ppo_rnd.sh will run the algorithms on the MountainCar environment.
This codebase is based on the excellent repo of Shangtong Zhang.