PingchuanMa / Respect-Learning

No Respect - No Learning

Respect-Learning

Setup

After following setup procedure on official website, run setup.sh for installing required packages.

Usage

Run run.py for local training and testing, run submit.py for online submission.

To-dos

Simulator

1. Understand observation, especially physical properties of joint
2. Figure out the reason why action repeat slows things down so much

Reward Shaping

1. Encourage bending knees
2. Activate muscle more for speed up
3. Encourage more aggressive stepping, currently the agent is too cautious to make a big step

RL Tricks

1. Discretize action value
2. Action noise & parameter noise
3. Layer normalization
4. Observation engineering

Benchmarks

1. Network structure
2. Activation function (SELU, RELU, ELU, ...)
3. 1D conv / FC