Super Auto Pets RL Agent

Gaming Using RL is my Final Year Project. This project is about building a Super Auto Pets Agent using sapai engine, sapai-gym open gym environment and stable baselines 3 for Reinforcement Learning Model. The Agent has 3 modules:

1. Computer Vision
2. Reinforcement Learning Model
3. Model2Game interface

1. Computer Vision

In this module, we extract state information(The Shop Pets and Food) from the game screenshot. This is done using Single Image Template Matching but the needle image and haystack image are of the same dimensions(we search the needle image in the haystack image). The implementation is in the image_detection.py

2. Reinforcement Learning Model

The RL Model is a MaskablePPO implementation using stable baselines 3. This model is run for 60,000 interations. It is a very weak model compared to humans. Its all time high wins is 2/10.

3. Model2Game interface

The Model2Game interface is used to convert the output of the model to mouse movements and clicks. This is implemented using pyautogui.

Training Metrics

Average Mean Reward

Average Number of Actions Per Game