SparceReward

This package provides comprehensive tools for examining rl algorithms on sparce and non-sparce environments.

At current version, we only offer Mujoco environment for tests. Current supported algorithms are as follows:

1- https://arxiv.org/pdf/1801.01290.pdf

2- https://arxiv.org/pdf/1706.01905.pdf

3- https://arxiv.org/pdf/1509.02971.pdf

4- ...

In the directory engine/reward_modifier, we provide different sparcity version for environments.

The typical code to run the program is as follows:

python3 main.py

If you want to change the environment you should type:

python3 main.py --env_name Ant-v2 The environments by defualt are non-sparce. You can make argument sparce by using the following command:

python3 main.py --env_name Ant-v2 --sparse_reward --threshold_sparcity 0.05 where --sparse_reward makes the environment's reward sparce and --threshold_sparcity determines the extent of sparcity.

If you want to change the algorithm to run:

python3 main.py --env_name --algo SAC

Defualt algorithm is DDPG. Current supported algorithms are: --algo SAC, --algo DDPG, --algo DDPG_PARAM

If you want to change the parameters of specific algorithm (for example SAC) you should write:

python3 main.py --env_name --algo SAC --tau_sac 1

which changes the tau parameter of SAC algorithm. Detailed information regarding the different setting of algorithms can be found in main.py argparser. We have provided detailed documentation for each parameter.