Title:
AdventumRL: A Quest-Based Reinforcement Learning API
AdventumRL: A Quest-Based Reinforcement Learning API
Author(s)
Singh, Kushagr
Advisor(s)
Riedl, Mark
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
We propose AdventumRL, a framework to facilitate complex, quest-based reinforcement learning in Minecraft, a 3D first person sandbox video game. We define a set of encodings on top of tools provided by Malmo, an open-source Minecraft reinforcement learning framework. We define a grammar framework for encodings states, actions, transitions, and goals that can represent quests in a reinforcement learning scenario. Proof of concept reinforcement learning agents are provided: a tabular Q-learning agent, utilizing a table to determine the best action to take from a given state, and two different deep Q- learning agents, which utilize a neural network instead of a table to determine the best action to take from a given state. One of the deep Q-learning agents solely utilizes the camera feed from Minecraft to determine its location, while the other directly uses positional and coordinate information instead. We demonstrate that the addition of our grammar framework allows the agents to complete a locked room quest that they could not complete without it.
Sponsor
Date Issued
2021-05
Extent
Resource Type
Text
Resource Subtype
Undergraduate Thesis