A simple project to try out some learning strategies on
Open AI Gym's CartPole-v0 environment :
- a solution based on genetic algorithm;
- a k nearest-neighbors strategy.
The idea is simplistic: make a fixed-sized population of agents compete. At every generation, survivors are selected stochastically according to their fitness (in this case their overall reward), and mutate into the next generation.
An Agent receives information about its environment (the observation variable
returned by the step() method of the environment), and makes a decision on the
action to take.
In our case, and Agent's brain consists of a simple two-layer
neural network, which is fed the observation output by the environment at the end of each step.
The genetic algorithm aims at determining the best parameters for the neural network.