MarioTennisAI

We generated a model of Mario tennis using a reinforcement learning algorithm called DQN. Mario runs on a reinforcement learning model, eventually grew to win Luigi.

Requirements

Tested PC spec is as below.

OS:Windows10, CPU:Core i5-2400, GPU:GTX-1060, Memory:16GB

Initial setting

1. The dolphin emulator. You can download from here. Stable version is better.
2. The Mario Power Tennis iso image.
3. Python 3. On Windows, you can use Anaconda.

Train

1. Launch MarioTennis on dolphin emulator.
2. Set windows looks lile the image.
3. Run the proggram.

cd ./src
python Tennis_AI_normal

Evaluation

You can see that as the number of games increases, the rewards for serving games also increase. However, in return games, the rewards are decreasing rather than increasing.

The breakdown of points is as above.

Play

Inference mode hasn't implemented, yet.

Movie

Commentary videos are uploaded to Youtube.

Article

http://hampen2929.com/mario-tenni-ai/825

This article explains in more detail in Japanese.