diplomsko delo
Žiga Kolar (Author), Igor Kononenko (Mentor), Petar Vračar (Co-mentor)

Abstract

Z napredkom strojnega učenja se je velik preboj naredil na področju športnega modeliranja. Podatkov play-by-play (podatki iz zapisnika) za eno sezono je več kot 400.000, kar je primerno za modeliranje košarkarske tekme z uporabo umetnih nevronskih mrež. Košarkarska tekma se modelira z dvema korakoma. Prvi je napoved naslednjega dogodka na tekmi, drugi pa je napoved časa med dvema dogodkoma. Algoritem prejme na vhodu vse prejšnje dogodke, kontekst tekme ter zmogljivosti ekip, vrne pa verjetnostno porazdelitev za naslednji dogodek. Slednja se nato vzorči in vrne naslednji dogodek na tekmi. Nato iz vseh dogodkov zgradimo verigo vseh dogodkov in dobimo vsak dogodek na tekmi, iz česar lahko izračunamo, kolikokrat se je vsak dogodek ponovil. Čas med dvema dogodkoma se modelira s homogenim modelom, ki na podlagi trenutnega časa vzorči vsa možna nadaljevanja. V naših poskusih se je za najboljši model izkazal model, ki za upodobitev ekip uporablja kombinacijo ratinga SRS in štirih faktorjev. Logiko vektorske vložitve smo preverjali z algoritmom TSNE. Ugotovili smo, da se algoritem nauči nekaj o ekipah, saj je venomer našel vsaj eno povezavo bodisi s številom točk, bodisi s katerim drugim atributom. Skladno s preteklimi študijami se je stavnica izkazala za najboljšo napovedovalko zmagovalcev košarkarskih tekem.

Keywords

globoko učenje;globoke nevronske mreže;modeliranje košarkarskih tekem;modeliranje play-by-play podatkov;računalništvo in informatika;univerzitetni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [Ž. Kolar]
UDC: 004.8:004.414.23(043.2)
COBISS: 59377155 Link will open in a new window
Views: 318
Downloads: 63
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Play-by-play modelling for basketball matches using vector embedding
Secondary abstract: In recent years there has been a big breakout in sport modelling due to the improvement of machine learning. There are more than 400.000 play-by-play data for one season, which is appropriate for modelling of basketball match with neural networks. Basketball match is modelled with two steps. First is the forecast of the next event, second is the forecast of time between two events. As input the algorithm receives all previous events, the match context and characteristics of the teams and returns the probability distribution for the next event. The latter is then sampled and returns the next event. Then we build a chain of events and we get every event of the match with which we can calculate how many times each event is repeated. Time between two events is modelled with homogeneous model. It samples the next time based on the current time. In our experiments the best model was the one that used rating SRS and four factors as input for team's characteristics. We checked the logic of the vector embedding with algorithm TSNE. We found out that the algorithm finds some knowledge for teams because it finds at least one connection with number of points or some other attribute. Based on previous research, the betting shop is the best for forecasting basketball victories.
Secondary keywords: deep learning;deep neural networks;basketball modelling;play-by-play data modelling;computer and information science;diploma;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000468
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 59 str.
ID: 12794555