<h2>AlphaZeroPlusPlus</h2>
<p>Objective: To train an RL system to play chess at ~1200 elo from scratch.</p>
<p>(How hard is it, really? Can it even be done on my home PC?)</p>
<p>Todo: Details and history.</p>
<h2>Previous attempts</h2>
<p>Previous attempts failed because of:</p>
<ul>
<li>Trying RL even before SL</li>
<li>Slow iteration speed</li>
<li>No tracking of progress</li>
<li>Not looking for help when stuck</li>
<li>No experimentation framework</li>
<li>No observability into the data/model</li>
<li>Not being in a presentable state</li>
<li>Trying a lot of things but not thinking mathematically</li>
<li>No Unit Tests! (Or any tests at all!)</li>
</ul>
<h2>New attempt</h2>
<p>We're gonna start from scratch again.</p>
<p>This time we will:</p>
<ul>
<li>Optimise iteration time
<ul>
<li>Start with small neural nets for fast training</li>
<li>Optimise algorithm for execution speed</li>
</ul>
</li>
<li>Do SL before RL</li>
<li>Track progress on this blog</li>
<li>Look up books/articles/papers etc for help when stuck</li>
<li>Develop an experimentation and observability framework</li>
<li>Visualise the models and data at every step</li>
<li>Spend some time making it presentable</li>
<li>Thinking more analytically about the problem before experimenting with code</li>
<li>Write unit tests!</li>
</ul>

1. The Start

AlphaZeroPlusPlus

Previous attempts

New attempt