Classification Challenge

I’ve been working on a game AI based around taking an input state and choosing one of 6 possible actions; in essence, a classifier with self-play. I’m having trouble achieving great accuracy, so I figured I’d open up a challenge.

This is just a personal project and I’m funding this myself, so I can’t go too crazy, but the prizes are:

• $256 to the top model
• $128 for second place
• $64 for third place

Data structure:

• Game stage as an integer (0-3)
• State values as 7 floats
• Correct action as an integer (0-5)

Rules:

• Must release the source to your model, under any license you choose
• Accuracy is measured against the testing set
• Don’t use the testing set for training
• Accuracy is measured by the minimum percentage of action accuracy (see example below)
• Competition will end Monday, 5/23 at noon EST
• Payout will occur by 5/26 via Paypal, Cash App, or Amazon gift card (participant’s choice)

Example accuracy:

• Action 0: 15000/31673 correct (47.35%)
• Action 1: 3000/5441 correct (55.13%)
• Action 2: 1000/2557 correct (39.1%)
• Action 3: 31/31 correct (100%)
• Action 4: 29/31 correct (93.54%)
• Action 5: 200/267 correct (74.9%)

Accuracy rating for this example: 39.1%

Dataset:

• Training
• Testing

Submit your entries by email to sera.brocious@gmail.com with the subject ‘Classification Challenge Entry’ and include both your accuracy and a link to a git repo (or otherwise) including the runnable source.

Good luck and happy hacking!

- Sera