valarzz / mpc_template-model_predictive_control_for_reinforcement_learning

Pytorch version of the MPC in model-based reinforcement learning (MBRL), currently only test in the CartPole-swing-up environment

MPC

This repo contains the implementation of Pytorch version of the MPC algorithm and the evaluation on the CartPole Swingup environment.

The MPC implementation is partly followed in this paper here and its repo

All the hyper-parameters and experiment setting are stored in the ./config folder.

All the results (figure and model) will be stored in the ./storage folder by default.

The training and testing scripts are stored in the ./scripts folder.

Requirements

• pytorch
• OpenAI gym

How to run

For the Cartpole swingup environment, simply go to the script folder and run

cd script && python train.py

The script will load the configurations in the ./config/config.yml file and begin to train.

By default, I only implemented the reward function for CartPole Swingup. Alternatively, you can train a NN-based reward model according to the needs of your environment. To do so, first run:

python train.py --train_reward_model 1

Then the reward model will be saved in the ./storage folder by default. Then you can run:

python train.py --use_reward_model 1

The MPC controller will use the learned reward function to perform optimizations.

Configuration explanation

In the config.yml file, there are 4 sets of configuration.

The model_config  part is the configuration of the parameters which determines the neural network architecture and the environment basis.

The training_config part is the configuration of the training process parameters.

The mpc_config part is the configuration of the MPC algorithm parameters, where you can choose to use the CEM optimizer or the Random optimizer.

If you want to train your model from scratch, then set the load_model parameter to False. If set to True, the trainer will load the model from model_path.