In our previous repository, we explored various models that could improve upon the baseline Multilayer Perceptron policy baseline and found that Recurrent Neural Networks work well. In this repository, we provide our code for our Recurrent Control Net, which beats Multilayer Perceptron and Structured Control Net baselines.

We provide scripts to design and train models. Upon exit, the weights of that training session are automatically saved. Below are run commands to render the environment with pretrained weights as well as log episodic reward during training.

To setup, see the CPG-RL repository for installing dependencies for MuJoCo.

# clone repository
git clone https://github.com/vliu15/RNN-CPG
cd RNN-CPG

Weights from training are automatically saved as .pkl files in weights.

# train on environments
python3 run.py --model rcn \
    --env HalfCheetah-v2 \
    --num_timesteps 2000000

In data collecting mode, the logs are written to .csv files in data.

# log training for plotting
python3 run.py --model rcn \
    --env HalfCheetah-v2 \
    --num_timesteps 2000000 \
    --collect_data

# render with pretrained weights
python3 run.py --model rcn \
    --env HalfCheetah-v2 \
    --num_timesteps 2000000 \
    --render \
    --weights_file /path/to/weights_file

Plots are written to .png files in plots.

# log episodic reward vs timestep, env is mandatory
python3 plot.py --env HalfCheetah-v2 \
    --avg_window 100

We continue to use MuJoCo v2 environments with OpenAI Gym as our means of testing our models. We find that MuJoCo provides a variety of different locomotive tasks that force a model to learn movements along different axes requiring different amounts of complexities.

We use Evolutionary Strategies as our optimization algorithm, as training through disturbance by random Gaussian noise has been proven to be very effective. We do not use another top alternative, OpenAI's Proximal Policy Optimization, because of its inflexibility to Recurrent Neural Networks (and extremely poor documentation).

Building off the idea of linear and nonlinear control modules outlined in this paper, the Recurrent Control Net uses a vanilla RNN as the nonlinear module and a simple linear mapping as the linear module. The RCN has:

• Nonlinear module:
  • 3 weight kernels
  • 2 bias vectors, should the n_use_bias flags be set to True
  • Layer activations after each kernel mapping
• Linear module:
  • 1 weight kernel
  • 1 bias vector, should the l_use_bias flag be set to True

We find that not using biases yields best results with Evolutionary Strategies as the training algorithm.

To cite this repository in publications:

@misc{RCN,
    contributors={Liu, Vincent and . Adeniji, Ademi and Lee, Nate and Zhao, Jason},
    title={Recurrent Control Nets as Central Pattern Generators},
    year={2018},
    publisher={GitHub},
    journal={GitHub repository},
    howpublished={\url{https://github.com/vliu15/RecurrentControlNets}},
}