Skip to content
/ ZSY2 Public

Training an Deep Q agent to learn ZSY. Properly, this time

Notifications You must be signed in to change notification settings

leonl0000/ZSY2

Repository files navigation

ZSY2

Training an Deep Q agent to learn ZSY.

争上游 (ZhengShangYou, or “Competition Upstream”) is a Chinese card game that is part strategy, part luck. Each player is dealt about 18 cards, they get rid of cards by matching patterns, and the player who gets rid of all their cards first wins.

Play the 2 player version of the game now! Tutorial level included for those new to ZSY. Game build with Unity.

Windows Build

Mac Build

In 2018, I attempted to create a Deep Q Agent to play this game better than humans, which you can read about here and here. To make a long story short, it failed by only having a win rate of ~40%. The agent was too simple, the training did not incorporate more advanced RL methods like fixed targets, and I didn't even have a good sample of humans to test it against.

This time, I seem to have reached human level performance. The 3 key improvements were: Conv nets, assembly models, and battle royale exploration. Read the final writeup here

About

Training an Deep Q agent to learn ZSY. Properly, this time

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published