Discussion Perfect SuperCoach

Joined
21 Apr 2018
Messages
19
Likes
94
#61
Answering my own question I just checked the weekly overall rankings for the GHOST team, they seem to have cracked the top ten 4 times (3 times before the bye)

Rd Pts Total Overall Rank
1 2413 2413 ??
2 2324 4737 ??
3 2322 7059 eq560th
4 2249 9308 eq89th
5 2196 11504 eq234th (lowest rank post the early rounds)
6 2369 13873 10th (9=13876)
7 2280 16153 25th (24=16155)
8 2324 18477 8th (7=18479)
9 2348 20825 3rd (2=20945) season high overall rank
10 2368 23193 22nd (21 = 23195)
11 2221 25414 10th (9th = 25421)

12 1467 26881 eq33rd
13 1731 28612 81st (eq79 = 28616) a big fall over the first 2 bye rounds
14 1992 30604 67th (66=30606)

15 2527 33131 20th(19=33134)
16 2519 35650 10th (9=35658)
17 2354 38004 30th (29=38008)
18 2365 40369 58th (57=40373)
19 2480 42849 52nd (51=42852)
20 2548 45397 65th (64=45404)
21 2299 47696 103rd (102=47699) big weekly fall
22 2754 50450 39th (37=50464) massive 1 week jump
23 2485 52935 eq77th another big weekly fall
This is amazing. Good work!
 

Beg2Differ

Rising Star Winner
Joined
22 Oct 2014
Messages
519
Likes
392
AFL Club
North Melb.
#62
Hi @IDIG,

1. How do you actually do it?

As I say in the article I use a branch of maths calling Mixed Integer Programming (MIP). I am trying to throw together an academic paper on it so I'll share that if I ever get around to it / if it ever gets published.

2. How long does it take you?

The optimal team one was reasonably straight forward so maybe a weekend to work out and code all of it up. I did that one last year and the rules didn't change so this year it took me maybe an hour to get everything set up again and then it took the computer about 12 hours to solve. The ghost team one was a lot trickier to model because of the emergencies so was thinking about it for a while but once I worked it out it took maybe 2-3 hours to code up and the computer 19 hours to solve. But yeah now I have it so next year I can just run it.

3. How did this all come about?

Like most obscure things, I bet a mate at the pub that I could work it out. Also came at a point during my PhD were nothing was working so needed to prove to myself that I could at least work this stuff out.

4. What else could this type of works be beneficial for in the 'real world' (ie. business, etc)

An example that might interest the users on this site is sport scheduling. So my colleagues at Monash are currently working with the Indian Premier League (IPL) to come up with the schedules that are fair on all the teams but also take into account stuff like ensuring players aren't away from home for too, popular matches are played during prime time, etc. My PhD is trying to apply it to schedule lab robotics more efficiently.

5. Can you put together next year's winning team for me and i'll split the $10k first prize with you?

That is essentially what happened with my brother. He got off to a ripper start so we decided to start plugging in the projections. The projections were pretty rubbish but still helped balance the side before byes and also recommended Angus Brayshaw pretty early on which was cool.


I am going to go with 9 :)
Hi Stevedwards

I am very impressed by your work here and am very interested in understanding more about it in regard to a specific applications I have grappled with for some time. I have sent you a private message in that regard and would appreciate any thoughts you may have.
 

chels

Rising Star Winner
Joined
8 Jan 2015
Messages
297
Likes
86
AFL Club
Sydney
#63
Hi @IDIG,

1. How do you actually do it?

As I say in the article I use a branch of maths calling Mixed Integer Programming (MIP). I am trying to throw together an academic paper on it so I'll share that if I ever get around to it / if it ever gets published.
Steve, I assume you have looked at the literature on cabin and flight crew rostering? It goes back to at least the '80's and (now) utilises MIP. Prof David Ryan at the University of Auckland's Department of Engineering Science has done excellent work both practical and published. He is a very approachable guy.

I always get a laugh when the afl tells us how dfficult it is to produce the fixture list. The constraints at (large) airlines are IMO somewhat more complex.
 
Joined
21 Apr 2018
Messages
19
Likes
94
#65
Steve, I assume you have looked at the literature on cabin and flight crew rostering? It goes back to at least the '80's and (now) utilises MIP. Prof David Ryan at the University of Auckland's Department of Engineering Science has done excellent work both practical and published. He is a very approachable guy.

I always get a laugh when the afl tells us how difficult it is to produce the fixture list. The constraints at (large) airlines are IMO somewhat more complex.
Not familiar with Prof David Ryan but am familiar with applying MIP to rostering and have seen some pretty amazing stuff being doing by the big airlines on that as well as flight planning. Will look into him. Yeah agree the fixture list for the AFL shouldn't be too hard to do fairly.
 

randomcliche

Rising Star Nominee
Joined
3 May 2017
Messages
231
Likes
245
AFL Club
Hawthorn
#66
This is amazing. Good work!
I second your thoughts, but about your work!

Hi Steve

I've got a question that I'd like to solve. 'What is the best starting structure?' If we could mix various phantom premiums, mid pricers, rookies with expected returns, would it be better to start a 0-22-8 or say a 13-0-17 type structure, or something in between like what I'm looking at 10-6-14 (premo-midP-rookie).

Def/Fwd premiums avg 110, mid ruck premiums 120, use the magic number (5428.64) for their prices.
Rookies, say average 60, a magic number of about 1.900 s.t price.
Mid-pricers, have say a magic number of say 3,333. (300kish)

It's a question I'd dearly love answered.

Regards
Andrew
 
Last edited:

IDIG

Moderator
Joined
8 Mar 2012
Messages
31,121
Likes
4,920
AFL Club
Essendon
#67
Hey @stevedwards

I know you generally do most of your amazing work after the season has finished but do you have any teasers for us to look forward to for this coming season?

Also, bump for those haven't seen this. Not sure it's the best thread to read as you work on your starting teams because it may sway you to being a little more clever and creative than you should :D
 
Joined
21 Apr 2018
Messages
19
Likes
94
#68
Hey @stevedwards

I know you generally do most of your amazing work after the season has finished but do you have any teasers for us to look forward to for this coming season?

Also, bump for those haven't seen this. Not sure it's the best thread to read as you work on your starting teams because it may sway you to being a little more clever and creative than you should :D
Hi IDIG,

Yes I should have an article coming out any day now - I wrote it about a week ago now so frustratingly it might be slightly out of date already given how quickly things change in the preseason.

It's got some pretty interesting ideas in it I think so will be very keen to hear the feedback from this forum. Hopefully with some help we can make it even better but will save that discussion til after the article is released.

Then should have a couple other articles out before the season starts so watch this space.
 
Joined
21 Apr 2018
Messages
19
Likes
94
#69
I second your thoughts, but about your work!

Hi Steve

I've got a question that I'd like to solve. 'What is the best starting structure?' If we could mix various phantom premiums, mid pricers, rookies with expected returns, would it be better to start a 0-22-8 or say a 13-0-17 type structure, or something in between like what I'm looking at 10-6-14 (premo-midP-rookie).

Def/Fwd premiums avg 110, mid ruck premiums 120, use the magic number (5428.64) for their prices.
Rookies, say average 60, a magic number of about 1.900 s.t price.
Mid-pricers, have say a magic number of say 3,333. (300kish)

It's a question I'd dearly love answered.

Regards
Andrew
Hi Andrew,

Sorry I completely missed this. I don't completely understand the question so just need to check a few things.

So is the question, given generic premium / mid / rookie options across all four lines, what is the best expected combination that you would be able to afford?

And then just estimating the scores and prices as something like follows?

1552261852701.png 1552261564789.png

Is that sort of what you mean?
 

randomcliche

Rising Star Nominee
Joined
3 May 2017
Messages
231
Likes
245
AFL Club
Hawthorn
#70
Hi Andrew,

Sorry I completely missed this. I don't completely understand the question so just need to check a few things.

So is the question, given generic premium / mid / rookie options across all four lines, what is the best expected combination that you would be able to afford?

And then just estimating the scores and prices as something like follows?

View attachment 7154 View attachment 7152

Is that sort of what you mean?
Yes, that's it with mid pricers averaging 90 and rookie prices of say $123,900k. Let's say 20 trade limit to account for some injury trades. Your insights would be most welcome.
 
Joined
21 Apr 2018
Messages
19
Likes
94
#71
Hey @IDIG.

https://www.heraldsun.com.au/sport/.../67dd2403ba5c49ab8e98a4baebc67f2e

Not sure if anyone saw but my article in the HS was published this week. It's essentially optimizing for percentage ownership to obtain the most cookie cutter team possible. I'm sure its not an original idea to use percentage ownership to select a starting team but actually doing it optimally while still respecting budget and positions is a little tricky.

The team as published in the article has imo too many rookies (18). I am thinking about constraining that to maybe 16 to force it to behave. Otherwise I was thinking about adding constraints to ensure dual position players to be scattered throughout the lines, i.e., one def/fwd in both defence and fwd. One ruck/fwd in both ruck and fwd, one def/mid in both def and mid and one med/fwd in both mid and forward... but maybe that is getting too cute.

I am also planning on entering the NRL supercoach competition with a similar strategy where I know absolutely nothing about NRL so any just general thoughts would be appreciated / is this doomed to fail?
 
Last edited:

stephen

400 Games Club
Joined
20 May 2014
Messages
2,169
Likes
836
AFL Club
St Kilda
#73
Hey @IDIG.

https://www.heraldsun.com.au/sport/.../67dd2403ba5c49ab8e98a4baebc67f2e

Not sure if anyone saw but my article in the HS was published this week. It's essentially optimizing for percentage ownership to obtain the most cookie cutter team possible. I'm sure its not an original idea to use percentage ownership to select a starting team but actually doing it optimally while still respecting budget and positions is a little tricky.

The team as published in the article has imo too many rookies (18). I am thinking about constraining that to maybe 16 to force it to behave. Otherwise I was thinking about adding constraints to ensure dual position players to be scattered throughout the lines, i.e., one def/fwd in both defence and fwd. One ruck/fwd in both ruck and fwd, one def/mid in both def and mid and one med/fwd in both mid and forward... but maybe that is getting too cute.

I am also planning on entering the NRL supercoach competition with a similar strategy where I know absolutely nothing about NRL so any just general thoughts would be appreciated / is this doomed to fail?
Thanks, Steve, this is so interesting. I think sheds light on the fact that a lot of SC players include too many expensive players, which forces too many rookies.

I'd be interested in how a team would look if it took the 15 most popular picks under 250k, then filled up with players above 10% popularity, starting from the cheapest up, but optimised to ensure all the money was spent.

(wish I could help with the NRL but know absolutely nothing and actually tried to put a team together using similar theory to above and the first game did not go well :/)
 

stephen

400 Games Club
Joined
20 May 2014
Messages
2,169
Likes
836
AFL Club
St Kilda
#75
Is embracing the wisdom of the crowds a smart move in SuperCoach?
Steven Edwards, Herald Sun

The “wisdom of crowds” is a concept at the junction of economics and psychology with the underlying principle that in some circumstances groups as a whole can often be smarter than the smartest people in the group.
The notion that the judgment of a crowd can be surprisingly accurate is generally traced back to an observation made by Charles Darwin’s half cousin Francis Galton in 1907. At a country fair, Galton observed that the average of the almost 800 entries in a “guess the weight of an ox” competition was alarmingly accurate – predicting the weight of the ox to the exact pound. Talk about a win for the people.

In SuperCoach it is possible to see how popular a player is by looking at their percentage ownership. It is generally an accepted strategy that you do not want to have too many players with a high percentage ownership otherwise there will likely be other teams that are very similar to yours and reduce the chance that you win overall.

This results in coaches searching for players with low percentage ownership who, in their opinion, are underestimated. These players give the team a point of difference and as such these types of players are referred to as PODs for short.

But what if instead of searching for PODs you were to do the exact opposite? Embrace the wisdom of the crowd and pick your starting squad to maximise percentage ownership. That is the starting strategy for my AI team, BOTempelli, this year. The squad below is the team that optimally maximises the percentage ownership based on data from March 2.

Anyone who has already been doing some research for the 2019 season will be quick to point out that this team is very generic. But that is exactly the goal. The team is in fact as generic as possible.

DEFENCE

The backline is stacked with Lloyd, Laird, and Whitfield as well as the popular mid-price option Zac Williams, who is returning at a discounted price after missing most of the 2018 season with an Achilles injury.

The rookie options also seem solid. Sam Collins is the 24-year-old ex-Fremantle player who was third in the 2018 VFL best-and-fairest. Isaac Quaynor, a solidly built under-18 All-Australian defender, won the goal kicking test at the 2018 Draft Combine with a perfect score. Marty Hore is a 23-year-old back-to-back best-and-fairest winner for Collingwood’s VFL side.

Unfortunately though, the crowd has selected Harrison Wigg, the injury riddled talent at the Suns who after repeat ankle injuries has a couple of weeks ago ruptured his ACL. If you have Wigg in your squad could you do me a favour and trade him out? Thanks.

MIDFIELD


Midfield premiums include Cripps, Fyfe, Kelly and Dusty Martin. Again, there is a popular mid-price option in Anthony Miles, a former 50-game player at Richmond and winner of the 2018 VFL best-and-fairest.

The midfield includes a very risky six rookies, including three from Carlton. Sam Walsh is the classy No.1 pick in the 2018 AFL draft. Liam Stocker won the best-and-fairest is last year’s TAC cup. Will Setterfield is a much hyped prospect who joins the Blues after missing the entire 2018 season at GWS with a ruptured ACL.

The remaining three rookies are Charlie Constable, a versatile 191cm running midfielder, Brett Bewley, a 23-year-old outside mid who was named in consecutive VFL teams of the year, and Nick Hind, a 24-year-old who should almost be picked purely based on his sprint down the middle of the ground in the final 10 minutes of the come from behind win in Essendon’s VFL semi-final last year. Although quite aggressive with so many rookies this seems like a reasonably strong midfield.

RUCKS

The rucks have both Grundy and Gawn, the two most expensive players in the whole competition not including the injured Tom Mitchell.

The rookie option in the rucks is Darcy Fort, a 25-year-old who quit his job as a civil engineer to return home to Geelong. However, it appears he is behind Rhys Stanley, Zac Smith, and Ryan Abbot, so Fort has a bit of work to do to break into the senior side.

FORWARDS

At this stage the forward line is somewhat terrifying, containing only two premiums and a total of six rookies. Dangerfield and Heeney are both very popular options this year and for good reason. After that though it becomes more challenging.

Three of the rookies appear to be solid candidates. Izak Rankine has been playing in the SANFL since he was 16 years old and was named in the under 18 All-Australian team twice. In addition, at the Draft Combine Rankine won the agility test and came third in the 20m sprint.

Ben Cavarra is a small forward who was in back-to-back VFL teams of the year and a prior Morrish Medalist as the best player in the TAC cup.

Matt Parker is an explosive forward from South Fremantle who, after enduring a supposedly troubled youth and bursting into tears upon hearing his name be called out at this year’s draft, looks hungry to pay back St Kilda for its faith in him.

The other three rookies that, in my opinion, have more uncertainty are Connor Rozee, pick 5 at the AFL draft, Chris Burgess, a 23-year-old contested marking defender, and Irving Mosquito, who is a very exciting prospect but at 18 years old will also be completing Year 12 and so it might be too much to expect him to consistently feature in Essendon’s senior side.

DUAL-POSITION PLAYERS
The crowd is currently selecting a very aggressive guns and rookie approach. The massive total of 18 rookies is making me a little bit nervous. The team is also somewhat rigid. In terms of dual-position players the only real flexibility comes from Setterfield in the midfield who could be swung forward in exchange for Dangerfield, Heeney or Mosquito.

It would also be nice if either Rozee or Burgess were selected in defence so that they could be moved between lines to cover holes. Hopefully when the crowd eventually trades out Wigg this might naturally correct itself. With so many rookies and such little flexibility I am currently concerned for BOTempelli’s prospects.

SHOULD THIS WORK?
So, will this “cookie cutter team” work? In his book The Wisdom of CrowdsJames Surowiecki proposes four criteria that separate wise crowds from irrational ones, three of which in my SuperCoach team appears to satisfy.

The first criteria is diversity of opinion – each person has their own interpretation of the facts. For example, I might think Patrick Cripps is God’s gift to football but you might mistakenly think he’s overrated.

The second criteria is decentralisation – people have access to draw on their own local knowledge. I believe this to be true for SuperCoach as everyone has that friend of a friend of friend who knows someone at the club who confidently claims that, for example, Bailey Scott will definitely be picked for North Melbourne in Round 1.

The third criteria is aggregation – there must exist some mechanism for turning private judgments into a collective decision. This is what BOTempelli aims at doing by using percentage ownership to determine the most “average” team.

The final criteria, and the criteria were I believe we might drop the ball so to speak, is independence – people’s opinions are not to be influenced by the opinions of those around them.

In SuperCoach there is so much information to stay on top of that often many coaches rely on experts in the forms of blogs, podcasts, and news articles like this one. This can result in somewhat of a herd mentality as many teams can be drawn to a small number of players.

Even writing this article is likely being counter-productive to the wisdom of the crowd by bringing everyone’s attention to a subset of the players, propagating their already high popularity. So I (very selfishly) encourage everyone this season to back their own judgment, do their own research and in doing so help improve the intelligence of the SuperCoach crowd and my chances of glory.

Steven Edwards is a PhD Candidate in the School of Mathematics at Monash University who works on large-scale optimisation problems.
 

freowho

Dual Best & Fairest
Joined
27 Jan 2014
Messages
3,636
Likes
997
AFL Club
Fremantle
#77
Bit disappointing actually but I'm assuming a lot of those names change before lockout. Will your team be updated during the rolling lockout Steven?
Maybe wisdom of the top 25% of the crowd!
 

stephen

400 Games Club
Joined
20 May 2014
Messages
2,169
Likes
836
AFL Club
St Kilda
#78
Thanks, Steve, this is so interesting. I think sheds light on the fact that a lot of SC players include too many expensive players, which forces too many rookies.

I'd be interested in how a team would look if it took the 15 most popular picks under 250k, then filled up with players above 10% popularity, starting from the cheapest up, but optimised to ensure all the money was spent.

(wish I could help with the NRL but know absolutely nothing and actually tried to put a team together using similar theory to above and the first game did not go well :/)
I realised that to do what I suggested was pretty simple and way below your paygrade so I had a crack at it. I had to change it to 16 rookies to make it work, and my only form of 'optimisation' was to just pick the most popular midpricer I could with the money I had left for the last pick (Libba). Team seems perfectly plausible to me. I reckon it would go alright.

unpod.JPG
 
Joined
12 Apr 2014
Messages
4,891
Likes
1,329
AFL Club
Collingwood
#79
I realised that to do what I suggested was pretty simple and way below your paygrade so I had a crack at it. I had to change it to 16 rookies to make it work, and my only form of 'optimisation' was to just pick the most popular midpricer I could with the money I had left for the last pick (Libba). Team seems perfectly plausible to me. I reckon it would go alright.

View attachment 7313
Nice work, and I agree that this side looks decent. If this was on RMT, the main observation would be that it's quite unbalanced across the def and forward lines ... but presumably this just indicates that there are some clear popular picks up forward and people are spreading the love down back, assuming I understand this correctly.

Edit: I just read Steven's (with a 'v') article, and it has almost the opposite structure. Maybe I was missing the effect of the price qualifier?
 
Last edited:
Joined
21 Apr 2018
Messages
19
Likes
94
#80
Bit disappointing actually but I'm assuming a lot of those names change before lockout. Will your team be updated during the rolling lockout Steven?
Maybe wisdom of the top 25% of the crowd!
Yeah the team was a bit out of date as the data was from March 2nd so essentially missed 2 full weeks of JLT. I have been having a play based on data from just now. I added some additional constraints to ensure that there are DPPs to swing between RUC/FWD, FWD/MID, DEF/FWD. I originally did DEF/MID as well but it kept selecting Andrew McGrath from Essendon and seemed like overkill. I also add a constraint to ensure no one worth over 200k has to start on the bench.

The new team looks pretty bloody good to me but you guys are the experts. Pretty cool that it selects duursma in the midfield anyway so I some reasonable flexibility between def/mid without constraining it.

1552709653236.png

Any thoughts? Has a whopping 165k left over. If I force it to have at most 50k left over I get the following instead it trades out Z.Williams for Scrimshaw and trades in Devon Smith for Balta so throws the balance out a little bit and only has 15.8k left over.

Hahaha this feels like cheating.
 
Top