Pole Balancing

Let a neural network learn how to balance poles that are mounted on a cart.

This is an example for a reinforcement learning problem. We will use direct policy search, i.e. we will not approximate a value function from which we will infer a policy. Instead, we will represent the policy with a neural network and we will optimize its parameters with a gradient-free optimization algorithm (CMAES).

You can choose between the environments SinglePoleBalancing and DoublePoleBalancing with or without velocities. The NeuroEvolutionAgent solves this problem.

Usage:

Available command line arguments are:

• "-spb": Single Pole Balancing
• "-po": Partially Observable (without velocities)
• "-ab": Use alpha beta filters to estimate the velocities
• "-des": Use double exponential smoothing to estimate the velocities

Start the experiment with the key "r". Increase the simulation speed with "+" and decrease the simulation speed with "-".