rd-tobias-sunderdiek / hyperparameter-tuning-demo

Demo for hyperparameter tuning with tune using an DQN example from Udacity for OpenAI gym

Hyperparametertuning Demo

This is a demo for hyperparameter-tuning with Tune[1] using an DQN example from Udacity[2] for the OpenAI-Gym environment[3] LunarLander[4].

This demo is meant to be able to be trained on cpu locally (took ~40 min. on a 2.5 GHz Quad-Core i7)

Goal

Land on the moon and get reward for landing properly, loose reward for using fuel or land outside landing pad. In this example, we use the metric mean_reward for this.

Install

• make install (tested with python 3.8)

Usage

• watch random, untrained agent via make random
• [optional] configure hyperparameter in train.py
• make train starts training
• see results in tensorboard via make tensorboard
• after training finished, make gif creates a .gif of the best model

[1] https://docs.ray.io/en/latest/tune.html

[2] https://github.com/udacity/deep-reinforcement-learning/tree/master/dqn/solution

[3] https://gym.openai.com/

[4] https://gym.openai.com/envs/LunarLander-v2/