An AI program that learns to play the 2003 GBA game Fire Emblem
I am playing the game on a downloaded emulator- visualboy advance and a downloaded ROM of Fire Emblem Blazing Sword. To play the game programmatically, the emulator must be in the top right corner of the screen, that is where the image processing software is looking to obtain data.
unit.py holds the Unit class, where each unit stores the information vital to completing the level. This information includes starting location, stats, Q-table with state-action rewards, appropriate dictionaries to map actions and states to the Q-table indices.
feaifuncs.py is a group of functions to ease the input commands into VisualBoyAdvance as well as capturing screens and doing light image processing and classification using machine learning methods.
prologue.py uses the baseline functions in feaifuncs to take turns in the prologue setting.
The Automated Playing notebook plays the game making random decisions, showing the functionality of the .py files.
The Prologue Q Learning notebook plays the game using a Q-table, updating values with rewards as the level progresses. Due to how slowly the game plays with the emulator, I instantitated the Q-table with distance-based rewards that would have been added step by step but can be calculated faster manually ahead of time. This saves us from having to fill the Q-table with values we already could calculate, saving runtime.
The Chapter 1 Solving notebook solves the second level using a greedy search with a Q-table that was instantiated with rewards rather than with 0's using an exploration-exploitation search algorithm. The search space is too large and the runs occur too slowly for a decaying exploration rate, so using a heuristic of 'Move toward the boss' starts our progress.