Generative prediction of basketball gameplay

Hafner, Andrej (author), Skočaj, Danijel (mentor)

Advancements in data analysis and machine learning have significantly impacted sports analytics, particularly in basketball, where detailed tracking data has enhanced understanding of player movements and game dynamics. However, existing models often focus on short-term trajectory predictions and omit key in-game events like shots, passes, and rebounds, limiting comprehensive game modeling. This thesis introduces a novel approach to generative modeling of basketball gameplay. By encoding games as sequences of game states through a basketball-specific grammar, we apply Transformer architectures to model gameplay similarly to language processing. We introduce Domain-Informed Decoding (DID) with output logits masking, integrating basketball rules directly into the decoding process to ensure coherent sequence generation. Additionally, we develop a method for quantifying generative realism by comparing key metrics between real and generated data. Our approach simulates entire basketball games, incorporating detailed player and ball movements with essential in-game events. Evaluations demonstrate that our model captures statistical properties of actual games and generates diverse, plausible gameplay scenarios, offering deeper insights into game mechanics and enhancing strategic planning in sports analytics.

• Language: English
• Keywords: multi-agent behaviour prediction, generative modeling, machine learning in sports
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-165148
• Publication date in RUL: 25.11.2024

Secondary language

• Language: Slovenian
• Title: Generativno napovedovanje poteka igre košarke

Abstract

Napredki v analizi podatkov in strojnem učenju so bistveno vplivali na športno analitiko, zlasti v košarki, kjer podrobni sledilni podatki omogočajo natančno razumevanje gibanja igralcev in dinamike igre. Obstoječi modeli se pogosto osredotočajo na kratkoročne napovedi gibanja in izpuščajo ključne dogodke v igri, kot so meti, podaje in skoki, kar omejuje celovito modeliranje igre. V tej magistrski nalogi predstavljamo nov pristop k generativnemu modeliranju košarkarske igre. Košarkarsko igro kodiramo kot zaporedje stanj igre s pomočjo košarki specifične gramatike, kar omogoča uporabo arhitekture Transformer za modeliranje igre na podoben način kot modeliranje jezika. Uvedemo domensko informirano dekodiranje (DID) z maskiranjem izhodnih vrednosti, ki vključi košarkarska pravila neposredno v proces dekodiranja za generiranje skladnih zaporedij žetonov. Poleg tega razvijemo metodo za vrednotenje generativnega realizma z primerjavo ključnih metrik med realnimi in generiranimi igrami. Naš pristop simulira celotne košarkarske tekme, kar vključuje podrobna gibanja igralcev in žoge s ključnimi dogodki v igri. Vrednotenje pokaže, da naš model zajame statistične lastnosti dejanskih iger in generira raznolike in verjetne scenarije, kar omogoča globlji vpogled v mehaniko igre ter izboljšuje strateško načrtovanje v športni analitiki.

• Keywords: napovedovanje večagentnega vedenja, generativno modeliranje, strojno učenje v športu