Cavescapes

A game that uses a Monte-Calro based learning algorithm to traverse a procedurally generated world.

Running the Game

Run MAIN.py using python 3 or above. Requires Numpy and PyGame installed.

Overview

As said in the heading, the Agent uses a monte-carlo based learning model. The world (maze) is created through the use of a modified reverse backtracking algorithm. Playable by a human or the AI.

• Use arrow keys to move and
• Turn on/off the AI by clicking the button on the top
• Speed up the learning process by clicking "fast learn". This increases the framerate and removes delay between frame updates
• The "change view" button allows you to toggle between seeing the whole board and just what the AI sees (It is able to see adjacent tiles only)
• The complexity/size of the grid can be changed by editing the grid_size variable in game_config.py
• For more advanced users, properties of the AI can be adjusted in game_config.py

Notable Behaviours and Quirks

• The AI seems to refuse to step on portal tiles, likely due to a high level of positive reinforcement towards regular path tiles that discourage the exploration of portals or new branches.
• The AI oftern retraces its paths to earn "slow and steady" rewards since a small positive reward is generally allocated to path tiles by the end of training.

Future Changes

Likely no major future changes.

-sonic597