Solving JumpIN’ Using Zero-Dependency Reinforcement Learning

Abstract

Reinforcement learning seeks to teach agents to solve problems using numerical rewards as feedback. This makes it possible to incentivize actions that maximize returns despite having no initial strategy or knowledge of their environment. We implement a zero-external-dependency Q-learning algorithm using Python to optimally solve the single-player game JumpIn' from SmartGames. We focus on interpretability of the model using Q-table parsing, and transferability to other games through a modular code structure. We observe rapid performance gains using our backtracking update algorithm.

Date
Feb 6, 2021
Event
Location
Virtual

Selected as finalist in Student Abstract category for a three-minute oral presentation in addition to the poster presentation. Based on project completed during summer 2020 supervised by Patrick Boily.

Rachel Ostic
Rachel Ostic
Data scientist

I’m fascinated by physics and how it combines abstract thinking with real world problems, and I am starting to fuse this interest with programming both inside and outside the lab.