MIER

Model Identification and Experience Relabelling

Installation

1. Run ./setup.sh

2. Add the following to ~/.bash_aliases, and then source ~/.bash_aliases

   alias mier='source activate mier ; export PYTHONPATH=<path to mier>; export MIER_DATA_PATH=<path to desired output directory>' After this, typing mier in any new terminal will set up the environment.

3. Mujoco installation

   • install mujoco in ~/.mujoco (we use mujoco 1.5)
   • add mujoco location and nvidia driver location to the LD_LIBRARY_PATH

Launchers

1. Meta-training :

   python launch_mier.py ./configs/envs/<env_name> ./configs/exps/train/<exp_name> --log_annotation <experiment name> --seed <seed>

2. Extrapolation :

   `python launch_mier.py ./configs/envs/<env_name> ./configs/exps/test/<exp_name> --log_annotation <experiment name> --load_model_itr <model iteration> --task_id_for_extrapolation <id> --seed <seed>`
   

exp_names for environments with only variable reward functions are mier-meta-train-only-rew.json (training) and mier-extrapolate-sep-models.json (extrapolation). For environments with variable dynamics, use mier-train.json (training )and mier-extrapolate.json (extrapolation). See .train_launcher.sh and extrapolation_launcher.sh for examples of how to launch experiments with gnu-parallel. The environment configuration file overrides the experiment configuration file. When running extrapolation, add value to load_path_prefix in the environment config file (see example in cheetah-negated-joints config file).