ARS is a random search method for training linear or fully connected neural network policies for continuous control problems, based on the paper "Simple random search provides a competitive approach to reinforcement learning."
Our ARS implementation relies on Python 3, OpenAI Gym, PyBullet and the Ray library for parallel computing.
To install OpenAI Gym follow the instructions here: https://github.com/openai/gym
To install PyBullet, use:
pip install pybullet
To install Ray execute:
pip install ray
For more information on Ray see http://ray.readthedocs.io/en/latest/.
Install this repo using pip or clone using git:
pip install arspb
First start Ray by executing a command of the following form:
ray start --head
or
ray start --head --redis-port=6379 --num-workers=18
This command starts multiple Python processes on one machine for parallel computations with Ray. Set "num_workers=X" for parallelizing ARS across X CPUs. For parallelzing ARS on a cluster follow the instructions here: http://ray.readthedocs.io/en/latest/using-ray-on-a-large-cluster.html.
We recommend using single threaded linear algebra computations by setting:
export MKL_NUM_THREADS=1
To train a policy for InvertedPendulumSwingupBulletEnv-v0, execute the following command:
python arspb/ars.py
All arguments passed into ARS are optional and can be modified to train other environments, use different hyperparameters, or use different random seeds. For example, to train a policy for InvertedPendulumSwingupBulletEnv-v0, execute the following command:
python arspb/ars.py --env_name InvertedPendulumSwingupBulletEnv-v0 --policy_type=linear --n_directions 230 --deltas_used 230 --step_size 0.02 --delta_std 0.0075 --n_workers 48 --shift 5
You can also train a fully connected neural network, specifying the sizes of the hidden layers, as follows:
python arspb/ars.py --env_name AntBulletEnv-v0 --policy_type=nn --policy_network_size=128,64 --n_directions 230 --deltas_used 230 --step_size 0.02 --delta_std 0.0075 --n_workers 48 --shift 5
By default, the activation function is tanh, you can also select clip, by adding this argument:
--activation=clip
First run a PyBullet GUI window using the following command:
python -m pybullet_utils.runServer
When running a gym environment, it will automatically connect to this GUI window over shared memory.
To render a trained policy, execute a command of the following form: (--render is not needed, since the env will connect to the running GUI server)
python arspb/run_policy.py --expert_policy_file=trained_policies/InvertedPendulumSwingupBulletEnv-v0/lin_policy_plus.npz --envname=InvertedPendulumSwingupBulletEnv-v0 --json_file=trained_policies/InvertedPendulumSwingupBulletEnv-v0/params.json
Or enjoy a fully connected neural network policy, AntBulletEnv-v0:
python arspb/run_policy.py --expert_policy_file=trained_policies/AntBulletEnv-v0/nn_policy_plus.npz --json_file=trained_policies/AntBulletEnv-v0/params.json
or a spinning running HumanoidBullet-v0 (click image below for video)
python3 arspb/run_policy.py --expert_policy_file=arspb/trained_policies/HumanoidBulletEnv-v0/nn_policy_plus.npz --json_file=arspb/trained_policies/HumanoidBulletEnv-v0/params.json --render
