Analysis Rolling Perfect Team

Joined
24 Mar 2015
Messages
1,630
Likes
2,980
AFL Club
North Melb.
#42
I tried to run the model for round 11 and it took about 16 hours to solve.....and then I realised I made a mistake by not including the best 18 constraint for rounds 10 & 11. Not sure how often I will be able to post this any more.
You've done some amazing work so far, it's been fascinating to see the evolution.

I'm wondering if you could set some heuristics (or do it manually) and exclude a whole stack of players from the analysis to make the model run faster? You may be doing something like that already or have played with the idea but it could reduce the search space a lot.
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#43
You've done some amazing work so far, it's been fascinating to see the evolution.

I'm wondering if you could set some heuristics (or do it manually) and exclude a whole stack of players from the analysis to make the model run faster? You may be doing something like that already or have played with the idea but it could reduce the search space a lot.
I suppose I could do some heuristics however first of all I would have to look up what that means.:geek:

I have resisted excluding players so far because working out exactly which ones I can exclude without affecting the result is quite tricky. However I have actually now done that and excluded the players yet to play at all which are not the 2 or three cheapest for each position. This reduced the variables from 50k to about 40k and should help.

The panacea for me was trying to shift the whole calc into a more powerful solver. I am using an a excel based solver called OpenSolver which is not that powerful however has been easy for me to pick up. I have been trying to work out how to do it in puLP via Solver Studio however the learning curve there is pretty steep. If anybody has any skills in that area and is keen to help then let me know.
 
Joined
15 Mar 2019
Messages
4,062
Likes
13,556
AFL Club
Hawthorn
#44
I suppose I could do some heuristics however first of all I would have to look up what that means.:geek:

I have resisted excluding players so far because working out exactly which ones I can exclude without affecting the result is quite tricky. However I have actually now done that and excluded the players yet to play at all which are not the 2 or three cheapest for each position. This reduced the variables from 50k to about 40k and should help.

The panacea for me was trying to shift the whole calc into a more powerful solver. I am using an a excel based solver called OpenSolver which is not that powerful however has been easy for me to pick up. I have been trying to work out how to do it in puLP via Solver Studio however the learning curve there is pretty steep. If anybody has any skills in that area and is keen to help then let me know.
If you want to cut it back a bit further, you could possibly exclude players that haven't scored at least 90-95 in a game yet. Looking at your last update Taylor has a top score of 100, with everyone else who has made the team having scored higher in at least one game.
 
Joined
24 Mar 2015
Messages
1,630
Likes
2,980
AFL Club
North Melb.
#45
If you want to cut it back a bit further, you could possibly exclude players that haven't scored at least 90-95 in a game yet. Looking at your last update Taylor has a top score of 100, with everyone else who has made the team having scored higher in at least one game.
This is the kind of thing I was thinking about.
If you look at it, having computed an optimal side for say 10 rounds, you can use that information to prune the search for what would be an optimal team after 11 rounds as it's not going to deviate too wildly.
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#46
So I excluded a lot of the non playing players as described above and ran it on a different PC. Solved in a couple of hours. Must have done something wrong in the first run (or there is something wrong in this one).

Total of 30,554 (avg 2774) compares to current leader of 24,663 (avg 2242).

1597660472205.png
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#47
Rolling perfect team after 12 rounds.

Total of 33,613 (avg 2801) compares to current leader at 27,143 (avg 2262).

Whilst it totally unrealistic that somebody could make all the right decisions to achieve a score anywhere near this it still staggers me that it is theoretically possible to achieve an average score of over 2800 per week particularly given there have been 3 best 18 rounds.

1597987527249.png
 
Last edited:
Joined
9 Dec 2015
Messages
1,046
Likes
538
AFL Club
Hawthorn
#49
Thanks mate.

I was wondering why the $7m team had done so well (900 pts in front of the current #1), partly due to the 5 perfect trades post Round 1 and partly because the $6m-$7m quote above was from memory when in fact $5m-$6m is the actual range. When I looked it up $5.6 million was the figure that would have allowed a perfect team to win the $50k in 2014.

If you try $5.5 million (instead of $7m) in a few weeks time I reckon that will be bang on as a second team that should finish close to #1 for the season but may lag a little behind early on noting a lot of the $400k plus players will be squeezed out by rookies or cheaper midpricers due to a further $50k per player reduction in the starting team.
Hey mate is there any chance you would be able to treat us to a post Round 13 edition of the $5.5 million salary cap Perfect SuperCoach side as well as the $10 million version?

It would be great to see how the cheaper team compares to the top real life SC sides in the overall rankings both now and at the end of the season.
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#50
I haven't run this for a while however I will be trying to run it for the full season after this round and this is a bit of a test run to see how long it would take. It took about 30 hours. Note this run only includes players that have played at least one game during the year so it may not be the theoretical max score.

This is the first run I have done where the maximum trade limit of 37 has applied so I was interested to see where those trades were saved. As it turns out it has only used 4 trades in round 2 and then used maximum for each round up until round 16 where it used 1 and then no trades in round 17.

This total score of 46918 (avg 2760) compares to the current leader of 38085 (avg 2240). Staggering that there is a 520 point difference between the averages.

1600490500489.png
 
Last edited:
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#51
I am working on the final perfect SuperCoach team. My first attempt finished this morning however I realised I made an error in one of the constraints so I have to rerun it. That run took 30 hours. I hope to be able to publish it on the weekend.
 
Joined
28 Aug 2019
Messages
15
Likes
34
AFL Club
Collingwood
#53
I've always wanted to learn how to do this, very interested on the mathematically best SC side every year, glad to see this is happening in the SC Scores community
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#54
After a run that look no less than 56 hours, in the immortal words of Paul Roos, "Here it is". The following shows the team that I believe would have achieved the highest possible score for Supercoach 2020. That score was 49,781 achieved at an average of 2,766 and compares to the winning score 40,466 (avg 2,248).

I would encourage everybody to have a close look at this and post if they see any anomolies. This is obviously quite a complex thing to do and takes at lot of time and computer grunt to achieve it. As you know I have posted a rolling team during the year and therefore the logic is reasonably well tested how it is not impossible that there could be flaws in the logic I have developed. I have checked the trade numbers and scoring players and other high level things but was keen to get it out there so I haven't exactly scoured the results to validate it. The movement of DPP would be the thing I am least confident about.

I would reiterate my acknowledgement of the role @stevedwards played in me be able to achieve this. Without the guidance he offered me in learning about Mixed Integer Programming (for a different reason initially) there is no way I would have been able to do this.

Finally I would also encourage people to post any quirky things they notice about this team and I will start with a couple.
  • You had to start Jack Viney, make him captain in the first round, and then trade him out.
  • You had to start Ben Cunnington and then trade him out for round 2.
1601179521990.png
 
Last edited:
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#55
Same team but presented in a different format which is hopefully a bit easier to see trades. Green means they are traded in that week and red means they were traded out that week. Note that DPPs do switch positions quite a bit.

1601184251651.png
 
Last edited:
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#56
Could a ghost ship have won the overall comp? By that I mean could a team selected for round 1 with captain, vice and emergencies set and no trades made during the year scored higher than the overall winner. I believe the answer to that question is yes.

The following team would have scored a total of 40,902 points at an average of 2,264 and beaten the overall winner by 436 points (total 40,466 at an average of 2,248).

Note that whilst I believe this would be very close to the highest scoring team of this nature I don't actually think it is the max score achievable. I have tried to calculate that and when I do it actually sets 2 emergencies on the mid and forward lines which become relevant during the bye rounds. However making the model select the lowest non-zero scoring emergency in the situation when only 1 is required has proven quite tricky so haven't quite got that to work yet.

1601260104951.png
 
Joined
1 Nov 2019
Messages
214
Likes
1,204
AFL Club
Richmond
#57
Could a ghost ship have won the overall comp? By that I mean could a team selected for round 1 with captain, vice and emergencies set and no trades made during the year scored higher than the overall winner. I believe the answer to that question is yes.

The following team would have scored a total of 40,902 points at an average of 2,264 and beaten the overall winner by 436 points (total 40,466 at an average of 2,248).

Note that whilst I believe this would be very close to the highest scoring team of this nature I don't actually think it is the max score achievable. I have tried to calculate that and when I do it actually sets 2 emergencies on the mid and forward lines which become relevant during the bye rounds. However making the model select the lowest non-zero scoring emergency in the situation when only 1 is required has proven quite tricky so haven't quite got that to work yet.

View attachment 22006
That's fantastic !

ok so next step is for me to find someone who had this team at the start and ruined it with transfers. just for the laugh...
 
Joined
22 Oct 2014
Messages
2,381
Likes
8,073
AFL Club
North Melb.
#58
That's fantastic !

ok so next step is for me to find someone who had this team at the start and ruined it with transfers. just for the laugh...
Actually what would be interesting is comparing what a team actually scored with what they would have scored had they left their round 1 team untouched for the season. How much is that agonising you do every round actually worth? I wonder if any teams would actually have score more by leaving it.

If you can extract those teams then I could work that out.
 
Joined
1 Nov 2019
Messages
214
Likes
1,204
AFL Club
Richmond
#59
Actually what would be interesting is comparing what a team actually scored with what they would have scored had they left their round 1 team untouched for the season. How much is that agonising you do every round actually worth? I wonder if any teams would actually have score more by leaving it.

If you can extract those teams then I could work that out.
so - which teams do you include in your sample ?
( I already have the composition of the first 100 teams for example)
 
Top