RL-asmus

• 

  2048 in python

• 

  human high score (while taking breaks from coding)

• 

  computer's high score (so far)

Inspiration

Wanted to try out reinforcement learning for a few years. Decided to apply this to https://play2048.co as a warm up problem

What it does

A neural network attempts to optimize it's high score at 2048

How I built it

Forked a university assignment codebase that had 2048 in python. Essentially rewrote the entire thing. Wrapped this in an OpenAI gym environment. Used a blog post that applies reinforcement learning to a different OpenAI gym.

Challenges I ran into

Original codebase was more confusing than helpful. Reinforcement learning seems to converge much slower than supervised learning.

Accomplishments that I'm proud of

Getting a clean codebase up. Catching bugs with unit tests. Will be proud if I can get good performance from the neural net.

What I learned

Training difficulty is nontrivial, even for such a simple game / state place.

What's next for RL-asmus

Try using RL to tune bicycle wheels (in real world)

Built With

• python
• pytorch