Please use this identifier to cite or link to this item: http://arks.princeton.edu/ark:/88435/dsp01pr76f653q
DC FieldValueLanguage
dc.contributor.authorBerman, Samuel "Sam"-
dc.date.accessioned2021-08-18T16:56:34Z-
dc.date.available2021-08-18T16:56:34Z-
dc.date.created2021-04-28-
dc.date.issued2021-08-18-
dc.identifier.urihttp://arks.princeton.edu/ark:/88435/dsp01pr76f653q-
dc.description.abstractThe objective of this project was to use reinforcement learning to create an AI that can play the video game Rocket League. In order to solve for the Bellman Equation of Rocket League using neural networks, a method of evaluating all possible future states in some time horizon had to found. This task was split up into two separate tasks, state prediction and state evaluation. The state prediction task was accomplished by estimating how the state of each of the cars and ball change through each time step. These estimations were made through a combination of data collection, mathematical approximations, physical intuition and estimations widely used by the RLBot community. The state evaluation task was accomplished training many different TensorFlow neural networks on both live data with rewards (reinforcement learning) and labeled state data from other Rocket League games (inverse reinforcement learning). All of the neural networks that were created failed to significantly distinguish between the evaluations of the states generated by the state prediction model. This is likely due to the fact that, despite being accurate, the state predictor is too slow. As a result, it was unable to predict enough states into the future to meaningfully distinguish between all of the projected states. Consequently, the neural networks could not properly evaluate which states were better because they were so similar. If reinforcement learning is to be successfully applied to Rocket League, it is almost certain that a different methodology will have to be implemented.en_US
dc.format.mimetypeapplication/pdf
dc.language.isoenen_US
dc.titleAn Exploration of Reinforcement Learning Through Rocket Leagueen_US
dc.typePrinceton University Senior Theses
pu.date.classyear2021en_US
pu.departmentMechanical and Aerospace Engineeringen_US
pu.pdf.coverpageSeniorThesisCoverPage
pu.contributor.authorid920192016
pu.mudd.walkinNoen_US
Appears in Collections:Mechanical and Aerospace Engineering, 1924-2021

Files in This Item:
File Description SizeFormat