Analysis AFL 2021: Rolling Perfect Team

Joined
22 Oct 2014
Messages
3,203
Likes
12,922
AFL Club
North Melb.
After a bit of trial and error I have discovered that if I introduce something called "branch and bound" tolerance to my score maximisation model it will solve much quicker. I don't know exactly what this means however assume it means that it isn't necessarily finding the absolute maximum score but something very close to it.

With that in mind the maximum score achievable up to round 15 (ie last week) was 40117 (avg 2674). This compares to the leader up to round 15 who had 33773 (avg 2252).


1625484240611.png
 
Joined
24 Mar 2015
Messages
2,333
Likes
6,278
AFL Club
North Melb.
Nice work! Just out of interest, have you compared R14 and R13 using branch and bound vs. normal for accuracy?

A 'pure' branch and bound that isn't doing any heuristics or estimation to determine bounds should still search the entire space, so will find the same solution (but hopefully faster!). In a problem space like this it should be able to discard a large number of search directions because they will be... 'crap' basically :p.
 
Joined
22 Oct 2014
Messages
3,203
Likes
12,922
AFL Club
North Melb.
Nice work! Just out of interest, have you compared R14 and R13 using branch and bound vs. normal for accuracy?

A 'pure' branch and bound that isn't doing any heuristics or estimation to determine bounds should still search the entire space, so will find the same solution (but hopefully faster!). In a problem space like this it should be able to discard a large number of search directions because they will be... 'crap' basically :p.
No I haven‘t tested previous rounds yet but intend to.

Note that the model I used for this solution was the one where I eliminated all player who had a max score of less than 80 except for those cheap DPPs (as outlined in the round 14 solution).
 
Joined
15 Mar 2019
Messages
6,354
Likes
23,247
AFL Club
Hawthorn
Looks like the popular Grundy and Gawn ruck line was definitely not the best way to go this year, at least to this point. No Gawn and holding off on Grundy until R15, never would have picked that.

No trades left for this team at this point - obviously things will change over coming weeks to optimise it, but I'm still expecting most trades to be used up early.
 
Joined
22 Oct 2014
Messages
3,203
Likes
12,922
AFL Club
North Melb.
Nice work! Just out of interest, have you compared R14 and R13 using branch and bound vs. normal for accuracy?

A 'pure' branch and bound that isn't doing any heuristics or estimation to determine bounds should still search the entire space, so will find the same solution (but hopefully faster!). In a problem space like this it should be able to discard a large number of search directions because they will be... 'crap' basically :p.
Tested the calculation for round 14.

Result without tolerance 37432
Result with 5% tolerance 37235

Difference of 197 points which is a little more than I would have expected.
 
Joined
22 Oct 2014
Messages
3,203
Likes
12,922
AFL Club
North Melb.
Maximum score (with 5% tolerance) achievable up to round 17 is 46,011 (avg 2,707). This compares to the leader up to round 17 who has 38,834 (avg 2,284).

Note the tolerance is included so the run completes in a reasonable amount of time. This took about 3 hours to solve. Based on the testing I have done I would expect this understates the true maximum by 200 to 300 points.

Note that this doesn't include round 18 as yet. I will run and post that in the next day or 2.

1626866888147.png
 
Joined
22 Oct 2014
Messages
3,203
Likes
12,922
AFL Club
North Melb.
I haven't done this for a while however this is potentially the most useful reference in knowing how to do well in this comp. The overall leader after round 18 is Supercoach Mumma with 41,279 points (avg 2,293). This is how she has got there.

1626956279594.png
 
Top