Hi,
Latest post is up where I continue to dive into what machine learning can discover. This time around I look at specific 5 man lineups, what they do well, not so well and apply thoughts on what coaches can do to get the most out of those lineups.
In past posts, I have looked at machine learning for high level scouting reports for teams. Now we dive into nitty-gritty of what specific lineups are doing and get more insightful findings.
http://www.zigzaganalytics.com/home/get ... e-learning
Getting the most out of NBA Lineups with machine learning
Re: Getting the most out of NBA Lineups with machine learnin
Watching the mavs vs heat game live...
Lineup 2 (my highest upside lineup for the mavs with Mejri ) just rattled off 8 straight points to get Dallas back in the game after a terribly slow start. 4 mins total for the lineup in the first quarter and they ended up a +2 after some late heat 3s.
The algorithm picked up that if this group is just average on offense they should end up with a positive +\- as Mejri really makes this group great on defence.
Lineup 2 (my highest upside lineup for the mavs with Mejri ) just rattled off 8 straight points to get Dallas back in the game after a terribly slow start. 4 mins total for the lineup in the first quarter and they ended up a +2 after some late heat 3s.
The algorithm picked up that if this group is just average on offense they should end up with a positive +\- as Mejri really makes this group great on defence.
Re: Getting the most out of NBA Lineups with machine learnin
The small sample sizes bug me. The coaching bugs me, not you.
Like the direction. Something similar for late playoff matchups would be good. Especially for the very top minute lineups. Though few of them will have much, enough testing. Grr.
Like the direction. Something similar for late playoff matchups would be good. Especially for the very top minute lineups. Though few of them will have much, enough testing. Grr.
Re: Getting the most out of NBA Lineups with machine learnin
In terms of sample size, lineup 2 with Mejri that I spoke about has played close to 80mins on the court together and has been involved in over 750 plays on the court this season. I have taken into account every 1 of those 750+ possessions, although I have summed together all the plays into a per game number for each lineup.
It is still a lot a lot of data if you look at how many possessions a lineup has been involved in.
I get that Dallas want to pay respect to dirk, it’s a hard situation, he’s a legend but for me it’s not like how the spurs were still playing Duncan 20-25mins a game when he was 40 as Duncan was still an elite defender. Dirks defence just isn’t up to scratch. Like I said it’s tough for Dallas, he carried them for over a decade, he deserves to go out playing and not riding the pine. Personally I’d be capping dirk at 20mins per game.
It is still a lot a lot of data if you look at how many possessions a lineup has been involved in.
I get that Dallas want to pay respect to dirk, it’s a hard situation, he’s a legend but for me it’s not like how the spurs were still playing Duncan 20-25mins a game when he was 40 as Duncan was still an elite defender. Dirks defence just isn’t up to scratch. Like I said it’s tough for Dallas, he carried them for over a decade, he deserves to go out playing and not riding the pine. Personally I’d be capping dirk at 20mins per game.
Re: Getting the most out of NBA Lineups with machine learnin
the spurs were still playing Duncan 20-25 mins a game when he was 40 as Duncan was still an elite defender
duncan ages 35-39 committed just 2.7 PF/40min...
mejri is currently committing fouls at a rate of 6.7 PF/40min...
his career rate (1613 min, ages 29-31) is 6.4 PF/40min...
at the age of 31 that rate won't be changing much, so he won't be playing significant minutes (say 25-30+ min/g) for anyone let alone dallas...
ages 29-31 mejri has averaged playing just 12 min/g. he could certainly play more, but the most anyone has played in that age range (past 40 years) committing fouls at a similar rate or worse were wayne cooper and alton lister (both 22 min/g), and players in that age range committing fouls at a similar rate have averaged playing just 17 min/g...
duncan ages 35-39 committed just 2.7 PF/40min...
mejri is currently committing fouls at a rate of 6.7 PF/40min...
his career rate (1613 min, ages 29-31) is 6.4 PF/40min...
at the age of 31 that rate won't be changing much, so he won't be playing significant minutes (say 25-30+ min/g) for anyone let alone dallas...
ages 29-31 mejri has averaged playing just 12 min/g. he could certainly play more, but the most anyone has played in that age range (past 40 years) committing fouls at a similar rate or worse were wayne cooper and alton lister (both 22 min/g), and players in that age range committing fouls at a similar rate have averaged playing just 17 min/g...
Re: Getting the most out of NBA Lineups with machine learnin
Leaguewide teams average 1.6 lineups that have been used 100 plus minutes for season and a total of 3.5 used over 50 minutes. Allow for injuries and matchup considerations that translates to less than 2 lineups used over 4 minutes per game in most games and about 2 more used over 2 minutes / gm. That leaves 50-75% of the game for most teams to be played by dink and superdink lineups.
If you have a good starting lineup, play it 15-20 minutes per game. Less than half the teams have played a lineup over 200 minutes or something like 8 minutes in 25 games. I don't see why you wouldn't test at least 5 other lineups 3-5 minutes per game for most games. I'd want at least that basis to evaluate them before playoffs.
If you have a good starting lineup, play it 15-20 minutes per game. Less than half the teams have played a lineup over 200 minutes or something like 8 minutes in 25 games. I don't see why you wouldn't test at least 5 other lineups 3-5 minutes per game for most games. I'd want at least that basis to evaluate them before playoffs.
Re: Getting the most out of NBA Lineups with machine learnin
Mejri for sure racks up fouls as he goes for a lot of blocks and contests everything HARD. He’s actually sneaky good at drawing charges, gets in really good position at times.
1 thing can be taught (although it’s breaking old habits) and that’s just using his “verticality” and having arms straight up in the air rather than going for so many blocks. I’ll refer to the spurs again, watch how they have Aldridge and gasol playing defence when right under the basket, it’s really simple but they literally put the arms straight up to avoid fouls, they really over emphasise it, and they still effect so many shots . Most spurs defenders play like this, it’s a big reason why the spurs are always leading the league in not comitting fouls.
So it can be done and taught.....but I agree it’s tough.
For me, I’d be playing Mejri 20 mins per game, right now he’s averaging 11....I’m certainly not saying he should play 30-35 a game.
I’d then have dirk playing about 20 with atleast half of those minutes next to Powell. Then give Powell 8 minutes at the 5 by himself.
Rough Split therefore would be Mejri 20.....dirk 20....Powell 18.
It gives dirk a nice send off still with solid minutes.
1 thing can be taught (although it’s breaking old habits) and that’s just using his “verticality” and having arms straight up in the air rather than going for so many blocks. I’ll refer to the spurs again, watch how they have Aldridge and gasol playing defence when right under the basket, it’s really simple but they literally put the arms straight up to avoid fouls, they really over emphasise it, and they still effect so many shots . Most spurs defenders play like this, it’s a big reason why the spurs are always leading the league in not comitting fouls.
So it can be done and taught.....but I agree it’s tough.
For me, I’d be playing Mejri 20 mins per game, right now he’s averaging 11....I’m certainly not saying he should play 30-35 a game.
I’d then have dirk playing about 20 with atleast half of those minutes next to Powell. Then give Powell 8 minutes at the 5 by himself.
Rough Split therefore would be Mejri 20.....dirk 20....Powell 18.
It gives dirk a nice send off still with solid minutes.
Re: Getting the most out of NBA Lineups with machine learnin
Mavs vs hawks live.....Lineup 4 (mavs starting unit with dirk at 5 and Kleber/barnes) weakness was opposition points in the paint according to the algorithm....first 8 points for Atlanta have come in the paint to start the game.
Re: Getting the most out of NBA Lineups with machine learnin
So by machine learning or otherwise, what does Carlisle tend to do when the opponent lights up inside and what is the logical Hawks response to that response and so on and so on...
Could you come up with a full lineup rotation that is a team's projected best response step by step to an opponent's typical rotation and have a set of variations for the most likely opponent lineup changes if they optimized to you or faced x number of common game scenarios? A plan and say 10-100-1000 variations. Sense where you are in game and go to optimal variation #35 or #921 or whatever and recalibrate response 5,10 or 30 times... Call it in from the tunnel or skybox or supercomputer to mobile phones... perhaps up to a million or billion potential matchup strategy "games" within a game... Do it by coaching experience / feel or supercomputer aided coaching experience / feel or just go with supercomputer experience / feel...
Could you come up with a full lineup rotation that is a team's projected best response step by step to an opponent's typical rotation and have a set of variations for the most likely opponent lineup changes if they optimized to you or faced x number of common game scenarios? A plan and say 10-100-1000 variations. Sense where you are in game and go to optimal variation #35 or #921 or whatever and recalibrate response 5,10 or 30 times... Call it in from the tunnel or skybox or supercomputer to mobile phones... perhaps up to a million or billion potential matchup strategy "games" within a game... Do it by coaching experience / feel or supercomputer aided coaching experience / feel or just go with supercomputer experience / feel...
Re: Getting the most out of NBA Lineups with machine learnin
Can you explain how there are 750 possessions or plays in 80 minutes?watto84 wrote:...played close to 80mins on the court together and has been involved in over 750 plays on the court this season. I have taken into account every 1 of those 750+ possessions,...
Counting plays by both teams, the Mavs get about [92.8 poss/48 * 80 * 2 = ] 309 poss. in 80 min.
Re: Getting the most out of NBA Lineups with machine learnin
“Plays” was the wrong wording on my part. I am using the play by play so hence called them “plays”. There are that many because every Action is a row in the play by play table. So going to the free throw line has a “foul” row and then 2 free throw rows for example. 750+ rows of data is probably better language to not confuse it with possessions