This repository contains a PIP package which is a modified version of the CartPole-v0 OpenAI environment which includes cart & pole friction and random sensor & actuator noise.

Installation

Install OpenAI gym.

Then install this package via

git clone http://github.com/AadityaPatanjali/gym-cartpolemod.git
cd gym-cartpolemod
sudo pip install -e .

Usage

Python usage:

import gym
import gym_cartpolemod

env = gym.make('CartPoleMod-v0')

Examples: Versions go from v0 through v6 for different noise scenarios

cd example_files
python deepQNetwork.py v1

The Environment

Some parameters for the cart-pole system:

• mass of the cart = 1.0
• mass of the pole = 0.1
• length of the pole = 0.5
• magnitude of the force = 10.0
• friction at the cart = 5e-4
• friction at the pole = 2e-6

Noise cases(v0 to v6):

• v0: Noise free
• v1: 5% Uniform Actuator noise
• v2: 10% Uniform Actuator noise
• v3: 5% Uniform Sensor noise
• v4: 10% Uniform Sensor noise
• v5: 0.1 var Gaussian Sensor noise
• v6: 0.2 var Gaussian Sensor noise

Note: The sensor noise is added to the angle, theta alone, and the actuator noise is added to the force.

Some Neural network parameters:

• Discount rate : gamma = 0.95
• Exploration rate : epsilon = 1.0
• Exploration decay : 90%
• Learning rate : 0.01
• NN Input layer : 24 neurons, tanh activation function
• NN Hidden layer : 48 neurons, tanh activation function
• NN Output layer : 2 neurons, linear activation function
• MSE loss function with Adam Optimizer
• Seed : None (Both in deepQNetwork and cartpolemod_env)

Note: To give a seed, uncomment line# 17 in deepQNetwork and change line# 76 in cartpole_mod to the desired seed value.

Notations:

• 1 run : 1000 trials
• 1 trials : 60,000 time-steps
• 1 time step : 0.02s
• Success rate : successful runs for 100 runs
• Successful trial: System lasts 60k time steps without failing
• Successful run : Atleast one successful trial in 1000 trials
• Average trials : Mean trials to successful trial over 100 runs
• Execution time : Approximate average time of execution for 100 runs

Results

CartpoleMod Version	Noise type	Success rate	Average trials	Execution time
v0	Noise free	100%	119.68	~ 4.5 hours
v1	5% Uniform Actuator Noise	100%	118.88	~ 4.5 hours
v2	10% Uniform Actuator Noise	100%	130.47	~ 4.5 hours
v3	5% Uniform Sensor Noise	100%	123.93	~ 5.5 hours
v4	10% Uniform Sensor Noise	100%	133.67	~ 5.5 hours
v5	σ^2 = 0.1 Gaussian Sensor Noise	100%	179.36	~ 6.5 hours
v6	σ^2 = 0.2 Gaussian Sensor Noise	98%	209.59	~ 8 hours

The team

• Aaditya Ravindran
• Apoorva Sonavani
• Rohith Krishna Gopi

This was created as a part of Prof. Jennie Si's class on Artificial Neural Computation (Fall 2017) at Arizona State University

Special thanks to MartinThoma, Kevin Frans, and keon