Mortal kombat 2 Sega genesis AI bot

Implementation of Mortal Kombat AI bot based on reinforcment learning. Deep Q-learning model is trained with gym-retro python package and tensorflow.

The neural network arhitecture

Conv Layer with 128 filter with kernel size 3 and input shape 4, 160 112
Conv Layer with 64 filter with kernel size 4
Global average pooling 2D
Dense layer 256
Dense layer 128
Dense with number of actions

Requirements

Python 3.6 requirements: -Gym-retro
-Gym
-Tensorflow 1.15
-Keras 1.15
-KerasRL (with some modifications)
-Baselines
-wandb

Files

Scripts

mortalkombat_env.py - setup of retro env for mortal kombat.
Implementaion of custom action and training wrapper for the env. train_model.py - training of deep q neural network
test_model.py - test for deep q neural network
train.sh - runes the the training script on slurm workload manager
test.sh - runes the the test script on slurm workload manager

Models

Trained models for Liu Kang.
14 models present