srkobakian
commented
5 years ago Goal: being able to pick Offensive verse Defensive states of players HMM: There are different states, HMMs guesses which state is currently happening (response) based on the data we have: ball trajectory, court geometry, or are able to create: ball angles Offensive: likely closer to the centre, in the box, can smash Defensive: further from the centre, closer to the base line
Hidden Markov Models: Two states, player will be in either one, state Offensive or Defensive We cannot directly observe the states. We can observe behaviour and attempt to split based on the behaviour at a given time period. We want the transition probabilities, assume you start in one, based on the behaviours (a function of the covariates which describe the current situation) are you going to start Offensive or move to a Defensive state.
We do observe other indicators: (check list on Jeremy's html Speed, eg. 60kmph v 30kmph observe different values Ratio of speed current shot/speed previous shot Server Score Number of shots in point (count) Return of first serve Return of second serve
Jeremy's work: HMMs used: distance from the net lags: angle of shot before, verse angle returning
Initial ideas: Indicators of ball movement: combine information based on variance of angles of previous hits in the rally Cross court shots: cumulative measure (count or proportion) to indicate how much players are moving around the court
General:
Find tutorials if you need help with: Github Rstudio
R skills course: https://dmac.netlify.com/ Textbook: https://r4ds.had.co.nz/