greenantoflw / pma

Predictable MDP Abstraction for Unsupervised Model-Based RL (ICML 2023)

Predictable MDP Abstraction for Unsupervised Model-Based RL

Project Page

Overview

This is the official implementation of Predictable MDP Abstraction (PMA). The implementation is based on the codebase of LiSP.

Installation

conda create --name pma python=3.8
conda activate pma
pip install -r requirements.txt

Examples

HalfCheeth (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env half_cheetah --dim_option 6 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 0.03 --tasks forward backward --penalty_lambdas 1

HalfCheetah (Classic model trained with random actions)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env half_cheetah --dim_option 6 --sample_latent_every 1 --z_eq_a 1 --collect_steps 4000 --tasks forward backward --penalty_lambdas 1

Ant (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env ant-v3 --plot_axis -40 40 -40 40 --dim_option 8 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 0.03 --tasks forward north --penalty_lambdas 20

Hopper (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env hopper-v3 --dim_option 3 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 50 --tasks forward hop --penalty_lambdas 1 5

Walker2d (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env walker2d-v3 --dim_option 6 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 5 --tasks forward backward --penalty_lambdas 1

InvertedPendulum (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env ip --dim_option 1 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 0.03 --tasks forward stay --penalty_lambdas 0 1 5

InvertedDoublePendulum (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env idp --dim_option 1 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 0.03 --tasks forward stay --penalty_lambdas 0 1 5

Reacher (PMA)

python main_pma.py --run_group Exp --memo X --max_path_length 200 --model_master_dim 512 --num_epochs_per_save 5000 --num_epochs_per_eval 1000 --num_epochs_per_log 50 --use_gpu 1 --seed 0 --fix_variance 1 --normalize_observations 1 --train_model_determ sepmod --ensemble_size 5 --replay_buffer_size 100000 --num_epochs 10002 --mppi_num_evals 2 --mppi_planning_horizon 15 --mppi_num_candidate_sequences 256 --mppi_refine_steps 10 --mppi_gamma 1.0 --mppi_action_std 1.0 --penalty_type disagreement --env reacher --dim_option 2 --sample_latent_every 1 --aux_reward_type disagreement --aux_reward_coef 0.03 --tasks default --penalty_lambdas 0 1 5

License

MIT