Abstract Demonstrations and Adaptive Exploration for Efficient and Stable Multi-step Sparse Reward Reinforcement Learning

This is the official repository for the codes of the A^2 paper.

Paper link: https://arxiv.org/abs/2207.09243

Contact: [email protected]

Installation

• Ubuntu 16.04 or Windows 10
• Clone and enter the repository, install requirements
  • git clone https://github.com/IanYangChina/TA2-paper-code
  • cd TA2-paper-code
  • python -m install -r requirements.txt
• Install Gym-minigrid
• Install Pybullet-multigoal-gym
• Install DRL_Implementation

Train

To train your own agent:

• Optionally conda activate YourCondaEnvironment
• python train.py --task gridworld_15 --agent dqn --render --num-seeds 2 --TA2 --eta 0.75 --tau 0.3
• This command means you will train a DQN agent on a gridworld task of size 15x15, with 75% demonstrated episodes and 0.3 adaptive exploration update speed for 2 random seeds in rendering mode.
• The --TA and --TA2 flags should not be used at the same time.

To see full argument explanation: python train.py -h, which gives:

usage: run.py [-h]
              [--task {gridworld_15,gridworld_25,block_stack,chest_push,chest_pick_and_place}]
              [--agent {dqn,sac,ddpg}] [--train] [--render]
              [--num-seeds {1,2,3,4,5,6,7,8,9,10}] [--TA] [--TA2]
              [--eta {0.25,0.5,0.75,1.0}] [--tau {0.1,0.3,0.5,0.7,0.9,1.0}]

optional arguments:
  -h, --help            show this help message and exit
  --task {gridworld_15,gridworld_25,block_stack,chest_push,chest_pick_and_place}
                        Name of the task, default: gridworld_15
  --agent {dqn,sac,ddpg}
                        Name of the agent, default: dqn
  --train               Whether to train or evaluate, default: False
  --render              Whether to render the task, default: False
  --num-seeds {1,2,3,4,5,6,7,8,9,10}
                        Number of seeds (runs), default: 1
  --TA                  Whether to use task decomposition & abstract
                        demonstrations, default: False
  --TA2                 Whether to use task decomposition, abstract
                        demonstrations & adaptive exploration, default: False
  --eta {0.25,0.5,0.75,1.0}
                        Proportion of demonstrated episodes, default: 0.75
  --tau {0.1,0.3,0.5,0.7,0.9,1.0}
                        Adaptive exploration update speed (a value of 1.0
                        means exact estimate instead of polyak), default: 0.3

Test a pretrained agent

To evaluate a pretrained agent:

• python test.py --task gridworld_15 --agent dqn --render --TA2
• This command means a DQN agent pretrained using TA2 on the gridworld 15x15 task will be evaluated in rendering mode.
• Each of the subgoals will be evaluated for 30 episodes.
• The --TA and --TA2 flags should not be used at the same time.

To see full argument explanation: python test.py -h, which gives:

usage: test.py [-h]
               [--task {gridworld_15,gridworld_25,block_stack,chest_push,chest_pick_and_place}]
               [--agent {dqn,sac,ddpg}] [--render] [--TA] [--TA2]

optional arguments:
  -h, --help            show this help message and exit
  --task {gridworld_15,gridworld_25,block_stack,chest_push,chest_pick_and_place}
                        Name of the task, default: gridworld_15
  --agent {dqn,sac,ddpg}
                        Name of the agent, default: dqn
  --render              Whether to render the task, default: False
  --TA                  Whether to use task decomposition & abstract
                        demonstrations, default: False
  --TA2                 Whether to use task decomposition, abstract
                        demonstrations & adaptive exploration, default: False

Citation

Acknowledgement