The autonomous gamer

a repository for the semestral project in the course Advanced Artificial Intelligence for Games on TEK SDU Odense

The project is focused on using AI to play simple games via screen-capture technology, drawing data directly from the screen and in real time. Main focus is on the chrome dino game, which, albeit simple, still makes the AI training hard by implementing a lot of randomness, each run having a different, generated track.

The technologies used are convolutional neural networks (pytorch backend), specifically ResNet and SqueezeNet and Deep Reinforcement Learning (see the dqn branch).

Prerequisites

The list of needed python libraries can be found in the file requirements.txt

List of usable scripts:

• dataCollect.py - collects data for the dataset, using screen capture technology
• dataCollect_flappy.py - collects data for the flappy bird dataset
• environment.py - mother class, containing utility functions for other scripts
• trainFromExisting.py - main training script for CNNs
• trainedAgent.py - script for testing pretrained models on real-time version of the game

In the DQN branch:

• main.py - runnable script, contains a CNN and an advanced DQN training script
• main_random.py - random agent testing script - used for benchmarking other models
• simple_main.py - basic DQN training script
• myDQN.py - contains utilities for DQN algorithms
• random_agent.py - contains utilities for the random agent
• replay.py - contains the algorithms simpleReplayBuffer and prioritizedExperienceReplayBuffer

Models

Pretrained models can be found in the models folder. The name of the model contains it's specifications: model_ architecture type _ number of layers _ number of epochs _ training dataset size

For example model_resnet-18_5_80k.pkl is a ResNet with 18 layers, trained on 80 000 images for 5 training cycles (epochs).