neverlmole / runai Goto Github PK

The code is about training a biped robot to walk. 

The codes is in 'environment\' folder. The main code is 'runner.py'. The front of 'runner.py' shows a simple usage of it. 'agent.py' and 'deepAgent.py' are the code for implementing different learning algorithm. 'game.py' and 'simulator.py' define the infrastructures. 'runner.xml' and 'runner2.xml' are two robot model in MuJoCo.
A sample training parameter file 'MCom_DeepQAgent_BasicGame.pam' is stored in 'data\' folder.  

Our code is based on MuJoCo, py_mujoco and pytorch, so you should install these three packages before trying to run our codes.
The install web page for py_mujoco is : 'https://github.com/openai/mujoco-py'

To see the result of the sample parameters, run the following command:

python runner.py -a DeepQAgent -f ../data/MCom --printrate 10000 --countrate 10000 -m Train -p {\'discount\':0.9999\,\'epsilon\':0.01\,\'epoch\':500\,\'netsize\':\[100\,30\,100\,30\]\,\'batchsize\'\:1000000\,\'numbatch\'\:5\,\'gpu\':True\,\'lr\'\:0.001\,\'p\'\:0.5} -g BasicGame --model ./runner.xml -d

To train a new deep Q-learing model, run the following command: 

python runner.py -a DeepQAgent -f ../data/new --printrate 10000 --countrate 10000 -m Train -p {\'discount\':0.9999\,\'epsilon\':0.5\,\'epoch\':500\,\'netsize\':\[100\,30\,100\,30\]\,\'batchsize\'\:1000\,\'numbatch\'\:5\,\'gpu\':True\,\'lr\'\:0.001\,\'p\'\:0.5} -g BasicGame --model ./runner.xml --renew -d

For discretized Q-learning, just change '-a DeepQAgent' to '-a QLearningAgent'. For discretized actor-critic, just change to '-a NaiveCrAcAgent'.