Winning card games with 1000+ CPUs

Vincent was playing a card game against his girlfriend and he kept loosing. So he wanted to train a bot to play on his behalf. This is our story.

We’re using AWS Lambda to get better at a card game named SushiGO. We make a small genetic algorithm in Python that uses AWS Lambda as a backend. The talk consists of these parts:

• Quick Explanation of the rules of the SushiGo Card Game
• Translation of real life to an algorithm
• Explain why this problem needs a lot of CPU
• Explain why AWS Lambda fits the simulation use-case
• How to quickly hack Concurrency in Python
• How to deploy lambda very quickly with chalice
• Experimentation Results

This talk will discuss an algorithm that we’ve tried to improve in three ways:

• Applying simple maths to make the search algorithm better
• Throwing lots (lots!) of CPU’s against the problem by leveraging AWS Lambda and python concurrency

We will conclude by discussing whether or not AWS Lambda is suitable for a gridsearch/grid simulation (hint, it’s not meant for this task, but it actually kind of works very well).