A keras solution for 2nd place NIPS RL 2017 challenge.

There is a slide, a lecture and a writeup(arxiv) about our work.

These instructions expect that opensim-rl conda environment is already setup as described in : https://github.com/stanfordnmbl/osim-rl/ .

$ source activate opensim-rl

Other dependencies is needed as follow

• Keras(since old version does not support selu activation)
• TensorFlow
• matplotlib
• numpy
• Pyro4
• parse
• pymsgbox(optional)

This version requires farming, before starting train.py, you should first start some farms by running python farm.py on each SLAVE machine you own. Then create a farmlist.py in the working directory (on the HOST machine) with the following content :

farmlist_base = [('127.0.0.1', 4), ('192.168.1.1', 8)]

# a farm of 4 cores is available on localhost, while a farm of 8 is available on another machine.

# expand the list if you have more machines.

# this file will be consumed by the host to find the slaves.

Try python farm.py --help to get more information about how to set the environment.

More information can be found in https://github.com/ctmakro/stanford-osrl .

Thanks to @ctmakro for providing us with this frame.

Test the model in parallel and calculate the average score.

We provide you with some trained parameters .

python test.py -a=10 -c=5 -t=200 -p logs

# test the model for 200 times with 10 actor networks and 5 critic networks ensemble

# the network parameters should be placed as logs/actormodel1.h5 ... logs/actormodel10.h5

Try python test.py --help to get more information .

• hzwer
• floz