Integrate Reinforcement Learning

[slavakurilyak]

Goal

As a developer, I want to develop and compare Reinforcement Learning (RL) algorithms, so that I can teach AI agents cryptocurrency trading.

Consider

• Analyzing work by others (here) to combine openai's gym and ccxt libraries

Inspiration

Learning to act optimally on time series input is of many practical uses in finance, healthcare, and industry. The value of taking an action depends on future actions and states, which makes it difficult to be modeled using a conventional supervised learning method. This is where (RL) fits.

-- Xiang Gao, 2018 (Research Paper, Source Code)

[bukosabino]

To explore: https://ai.googleblog.com/2018/08/introducing-new-framework-for-flexible.html

[slavakurilyak]

In order to understand the potential of reinforcement learning, it is important to look at the AlphaGo Zero case study.

AlphaGo Zero is the strongest Go player in the world. It outperformed all previous versions of AlphaGo. It defeated the version of AlphaGo that won against the world champion Lee Sedol by 100 games to 0.

What was the difference between AlphaGo and AlphaGo Zero? AlphaGo was trained by supervised learning from human expert moves, and by reinforcement learning from self-play. AlphaGo Zero was trained solely on reinforcement learning, without human data.

If you want to learn more about AlphaGo Zero, watch this video (2 min).

[slavakurilyak]

DeepMind releases trfl, a reinforcement learning framework built on top of tensorflow. https://github.com/deepmind/trfl/blob/master/docs/index.md