https://github.com/dmitrijsc/practical-rl
Reinforcement learning in Julia. Solving OpenAI gym.
https://github.com/dmitrijsc/practical-rl
cross-entropy frozenlake julia julialang mxnet openai openai-gym q-learning reinforcement-learning
Last synced: 6 months ago
JSON representation
Reinforcement learning in Julia. Solving OpenAI gym.
- Host: GitHub
- URL: https://github.com/dmitrijsc/practical-rl
- Owner: dmitrijsc
- License: mit
- Created: 2017-12-07T18:24:07.000Z (almost 8 years ago)
- Default Branch: master
- Last Pushed: 2018-03-14T22:55:25.000Z (over 7 years ago)
- Last Synced: 2025-03-24T09:53:00.326Z (7 months ago)
- Topics: cross-entropy, frozenlake, julia, julialang, mxnet, openai, openai-gym, q-learning, reinforcement-learning
- Language: Julia
- Homepage: https://solveai.net
- Size: 45.9 KB
- Stars: 43
- Watchers: 3
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE.md
Awesome Lists containing this project
README
# Reinforcement learning in Julia
Purpose of this project/ repository is to replicate [Practical RL course](https://github.com/yandexdataschool/Practical_RL) exercises in Julia.
About Practical RL (from their GitHub): A course on reinforcement learning in the wild. Taught on-campus in HSE and Yandex SDA.
## Setup
This code was tested on Julia 0.6.1. It will be using MXNet for all deep learning activities.
Please follow standard process to configure `Open AI gym`, `POMDPs.jl` and `MXNet.jl` from the corresponding package repository.
## Table of contents
The following objectives are implemented and available as a separate Julia files:
week_0: FrozenLake with genetic algorithms
- FrozenLake (4x4): average score 0.86 ([post](https://solveai.net/2017/12/07/playing-frozenlake-with-genetic-algorithms/))
- FrozenLake (8x8): average score 0.97
week_1: Cross-Entropy Method
- FrozenLake8x8, Taxi-v2 (CEM) ([post](https://solveai.net/2017/12/24/playing-frozenlake-using-cross-entropy-method/), [post2](https://solveai.net/2017/12/24/difference-between-evolutionary-methods-and-methods-that-learn-value-functions/))
- CartPole-v0 (Deep Cross-Entropy Method using MXNet): average score 200.0 ([post](https://solveai.net/2018/01/06/playing-cartpole-with-deep-cross-entropy-method-using-julia-and-mxnet/), [post 2](https://solveai.net/2018/01/08/importance-of-learning-rate-when-running-deep-cross-entropy-method/))
week_2: Q-learning (Value-table method)
- Taxi-v2: average score 8.7/8.5 ([post](https://solveai.net/2018/01/15/julia-q-learning-using-value-table-to-solve-taxi-v2/), [post 2](https://solveai.net/2018/02/01/julia-q-learning-and-epsilon-discount-factor/))
week_3: SARSA
- Taxi-v2: 8.5/8.5 ([post](https://solveai.net/2018/02/01/julia-solving-open-ai-taxi-v2-using-sarsa-algorithm/))
week_4: DQN
- CartPole-v0: 100/100 in 1 episode ([post](https://solveai.net/2018/02/27/playing-cartpole-using-julia-and-mxnet-implementation-of-dqn/))
- Space Invaders: ~220 in average ([post](https://solveai.net/2018/03/12/playing-space-invader-using-dqn-julia-and-mxnet/))
## T&C
I will try to follow Julia best practices when writing code and optimize it whenever and wherever possible.
Please excuse in case of any serious issues and you are welcome to submit your PR.
## Contacts
Feel free to contact me over Issues, solveai.net or Julia Slack channel @dmitrijsc.