This repository contains a PIP package which is an OpenAI environment for simulating an enironment in which bananas get sold.
Installation
Install the OpenAI gym.
Then install this package via
pip install -e .
Usage
import gym
import gym_banana
env = gym.make('Banana-v0')
See https://github.com/matthiasplappert/keras-rl/tree/master/examples for some examples.
The Environment
Imagine you are selling bananas. One at a time. And the bananas get bad pretty quickly. Let's say in 3 days. The probability that I will sell the banana is given by
where x-1 is my profit. This x-1 is my reward. If I don't sell the banana, the agent gets a reward of -1 (the price of the banana).
Example agent
See .
Analytic solution
This recursive equation corresponds to the probability tree diagram. Prices on the n-th day are x_n.
Optimization in all x_i (eg. via a grid) allows to compare (at low N) the learned result with the analytic result (except for statistical fluctuations).
See .
Comparison
The mean reward is roughly the same as the analytic for N=2:
[1.5 1.1] => -0.25655004382636204
This means highest reward on average is achieved by selling for 1.5 on the first day and 1.1 on the second.