Value Consistency Prioritization (VCP)

Code for "Efficient Multi-Goal Reinforcement Learning via Value Consistency Prioritization".

Requirements

python3.6+, tensorflow, gym, mujoco

Installation

• Clone the repo and cd into it:

  git clone https://github.com/jiawei415/VCP.git
  cd VCP

• Install vcp package

  pip install -e .

Usage

Environments: PointMassEmptyEnv-v1, Reacher-v2, FetchReach-v1, HandReach-v0, HandManipulatePenRotate-v0.

VCP:

python -m  vcp.run --env PointMassEmptyEnv-v1 --num_epoch 50 --num_env 1  --alg_config "{'k_heads':16,'priority_temperature':9.0}

HER:

python -m  vcp.run --env PointMassEmptyEnv-v1 --num_epoch 50 --num_env 16  --alg_config "{'k_heads':1,'prioritized_replay':False,'use_her_buffer':False}"