<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<title>Frozen Lake</title>
<meta name="viewport" content="width=device-width, initial-scale=1">
<style>
body {
background-color: #000;
}
#container {
margin: auto;
max-width: 800px;
text-align: center;
}
#container>img {
width: 100%
}
#container>a,
#container>h2,
#container>p {
color: #fff;
}
#container>a {
margin-top: 16px;
}
</style>
</head>
<body>
<div id="container">
<h2>RL - Slippery Frozen Lake Q-Learning</h2>
<p>I trained a Q-Learning model on the OpenAI Gym Slippery Frozen Lake environment for 20,000 iterations, and
evaluated for 1,000 iterations.
The trained model had a success rate of about 73%. Action for the 3 column on the second row is especially
interesting,
it correctly learns that the best policy is to try move toward one of the holes because there is a 1/3
chance of slipping and slips are orthogonal to the desired direction.</p>
<img src="assets/eval_screenshot_train_20k.png" alt="Q-Learning Agent">
<a href="https://www.youtube.com/watch?v=b1oh3TK6Jhg">Training and Evaluation Video (youtube)</a>
</div>
</body>
</html>