Analysis 2023 AFL SC: Rolling Perfect Team [Post #242 onwards]

[Beg2Differ]

Maximum score for the first 3 rounds is 4399 (avg 1466) which compares to the current leader who has 3181 (avg 1060).

 

[Beg2Differ]

This is what the current leader has done. Fair to say he has nailed his captain so far.

 

[Diabolical]

This is what the current leader has done. Fair to say he has nailed his captain so far.

View attachment 36657

Smart play once nailing Henriques and doubling up with Short to move to the safety of Maxy. I reckon if I nailed Henriques in R1 I would have probably taken my luck and gone Rashid in R2.

 

[Beg2Differ]

Smart play once nailing Henriques and doubling up with Short to move to the safety of Maxy. I reckon if I nailed Henriques in R1 I would have probably taken my luck and gone Rashid in R2.

I suppose that is probably what you have to do to win it all. I doubt I would ever be brave enough.

 

[Herbie66]

This is what the current leader has done. Fair to say he has nailed his captain so far.

View attachment 36658

Traded in Couch , DreRuss & Qais.

Form over reputation bypassing all 3 of NCN , Stoinis & Zampa

Huge call.

 

[Beg2Differ]

Finally I have had some success in determining the max score achievable in AFL Supercoach for 2021. The following team is the max achievable with a 5% tolerance. That means it won't be the absolute highest but it will be a pretty decent approximation. To illustrate the point I can see that the model hasn't selected the highest scoring player as captain in round 2. Other than that it kind of passes the sniff test.

This team would have scored 62,839 points at an average of 2,732. This compares to the 2021 winner Supercoach Mumma who scored 53,859 at an average of 2,342.

 

[Beg2Differ]

After many months, many attempts and many #*?^$*% I believe I have successfully determined the highest AFL supercoach score achievable for season 2021. The one little rider is that I did manually remove a number of players from the model who I felt had no chance of being in the perfect team based on the combination of their games played, price and scores.

The max score 63,223 (avg 2,749) compares to the 5% tolerance score (above) of 62,839 (avg 2,732) and the winning score of 53,859 (avg 2,342).

However the biggest lesson out of this is to run the model on my work computer. Whilst it took 62 hours to solve I did run a similar model on my personnel computer just after the end of the season for 3 days (72 hours) without it finishing.

 

[Beg2Differ]

Maximum score for the first 4 rounds is 5801 (avg 1450) which compares to the current leader who has 4275 (avg 1069).

 

[Beg2Differ]

The leader held is spot in round 4.

 

[mike89]

Leader has absolutely nailed every captain so far. That's absurd. Especially considering who the first two were!

 

[Ironhawk]

Leader has absolutely nailed every captain so far. That's absurd. Especially considering who the first two were!

Yep. I've managed the best captain in R3 and R4 (not all that difficult with who they were) and thought that was a good effort. 4 in a row is impressive!

 

[warewolves]

After many months, many attempts and many #*?^$*% I believe I have successfully determined the highest AFL supercoach score achievable for season 2021. The one little rider is that I did manually remove a number of players from the model who I felt had no chance of being in the perfect team based on the combination of their games played, price and scores.

The max score 63,223 (avg 2,749) compares to the 5% tolerance score (above) of 62,839 (avg 2,732) and the winning score of 53,859 (avg 2,342).

However the biggest lesson out of this is to run the model on my work computer. Whilst it took 62 hours to solve I did run a similar model on my personnel computer just after the end of the season for 3 days (72 hours) without it finishing.

View attachment 36760

Great work @Beg2Differ – it must have been a really programming challenge to deal with all of those SC parameters/constraints.
I always take an interest in these retrospectives – hoping (in vain so far) for some insights on the best way to set up for SC and Fantasy. This inevitably comes around to the merits of the Guns & Rookies approach – which is almost always promoted as the best way to go.
I am in favour a value-based (ie mid-pricer) set-up and am looking at this analysis as some sort of vindication on that front. Using a somewhat generous definition of over $500K for Guns, the attachment is my breakdown of the optimal 2021 starting squad.
This is really nothing lile a G&R approach which relies on about 13-14 Guns & the rest nearly all rookies. Of course nobody could ever dream of getting close to the optimal score, but it stands to reason (imo) that the path to the optimal score is the value-based MP, then picking your team in that manner is at least OK and at best, the preferred approach.
The optimal starting team also casts doubt on the set & forget rucks orthodoxy too – going for value is the key. BTW I’m not suggesting its easy to pick the eyes out of the mid-pricers – but the principle is valid.
As an example if you look at Dunkley as a case study. He was a Mid-Fwd DPP and you’d think for sure he’d be in the starting Fwds, but no, it was best to start him as a Mid – presumably because the model wanted to fit in another value-packed Fwd.
Having worked on this a lot more than the rest of us, I would be interested on your thoughts @Beg2Differ

 

[Ironhawk]

Great work @Beg2Differ – it must have been a really programming challenge to deal with all of those SC parameters/constraints.
I always take an interest in these retrospectives – hoping (in vain so far) for some insights on the best way to set up for SC and Fantasy. This inevitably comes around to the merits of the Guns & Rookies approach – which is almost always promoted as the best way to go.
I am in favour a value-based (ie mid-pricer) set-up and am looking at this analysis as some sort of vindication on that front. Using a somewhat generous definition of over $500K for Guns, the attachment is my breakdown of the optimal 2021 starting squad.
This is really nothing lile a G&R approach which relies on about 13-14 Guns & the rest nearly all rookies. Of course nobody could ever dream of getting close to the optimal score, but it stands to reason (imo) that the path to the optimal score is the value-based MP, then picking your team in that manner is at least OK and at best, the preferred approach.
The optimal starting team also casts doubt on the set & forget rucks orthodoxy too – going for value is the key. BTW I’m not suggesting its easy to pick the eyes out of the mid-pricers – but the principle is valid.
As an example if you look at Dunkley as a case study. He was a Mid-Fwd DPP and you’d think for sure he’d be in the starting Fwds, but no, it was best to start him as a Mid – presumably because the model wanted to fit in another value-packed Fwd.
Having worked on this a lot more than the rest of us, I would be interested on your thoughts @Beg2Differ

I started Dunkley as a mid. I did get a lot of starting selections wrong, but thought it was the best option with him: there seemed enough value forward options I thought could score comparably to mids, I just didn't choose the best ones. Think I had 9/30 of the best starting team, with only Butters, Ziebell and Impey in the forwards.

Forwards look OK value again from what I have seen, but I don't have access to all the prices. Seems there will possibly be more top end forward options than 2021, so I'm not sure you will want to go the same way in 2022 - possibly one or two higher priced M/F should be started as mids again?

 

[Beg2Differ]

Maximum score for the first 5 rounds is 6925 (avg 1385) which compares to the current leader who has 4761 (avg 952).

 

[Beg2Differ]

New leader this week (12th man) who has scored 4761 at an average of 952.

Former leader (Cricket) scored 469 in round 5 to drop to 3rd.

 

[Beg2Differ]

Great work @Beg2Differ – it must have been a really programming challenge to deal with all of those SC parameters/constraints.
I always take an interest in these retrospectives – hoping (in vain so far) for some insights on the best way to set up for SC and Fantasy. This inevitably comes around to the merits of the Guns & Rookies approach – which is almost always promoted as the best way to go.
I am in favour a value-based (ie mid-pricer) set-up and am looking at this analysis as some sort of vindication on that front. Using a somewhat generous definition of over $500K for Guns, the attachment is my breakdown of the optimal 2021 starting squad.
This is really nothing lile a G&R approach which relies on about 13-14 Guns & the rest nearly all rookies. Of course nobody could ever dream of getting close to the optimal score, but it stands to reason (imo) that the path to the optimal score is the value-based MP, then picking your team in that manner is at least OK and at best, the preferred approach.
The optimal starting team also casts doubt on the set & forget rucks orthodoxy too – going for value is the key. BTW I’m not suggesting its easy to pick the eyes out of the mid-pricers – but the principle is valid.
As an example if you look at Dunkley as a case study. He was a Mid-Fwd DPP and you’d think for sure he’d be in the starting Fwds, but no, it was best to start him as a Mid – presumably because the model wanted to fit in another value-packed Fwd.
Having worked on this a lot more than the rest of us, I would be interested on your thoughts @Beg2Differ

Thanks for taking an interest.

Yes it has been a fair challenge to get to this point however I have been able to build my knowledge over the last few years and feel as though I am mostly on top of it now. The biggest barrier now is the software I have access to. Generally my ambition now tests the limits of that software and getting higher powered software would come at a fair cost both in terms of dollars and another learning curve.

Anyway back to your question. My own personal view is that whilst I find this an interesting academic exercise I don't think it offers too many strategic insights into how to succeed at this game. Generally what you see is that it trades players that are about to have a lull (either form, role or injury) for players that are about to go on a strong run. In my view that is impossible to predict and anybody attempting to do so will fail.

The problem with the value based approach is that relies on you getting the vast majority of value selections correct and history tells us that is nigh on impossible. If your success rate isn't high enough then you consign yourself to a miserable season because you will slip off the pace and there aren't enough trades available to correct the situation, be it switch to G&R or different value selections. I don't really think that the max score scenario supports that a value based approach can work because it so extreme and therefore unrealistic; however equally I don't necessarily think it means it can't work either.

In order to truly assess the merits of each approach I think some type of probabilistic modelling approach which builds in realistic decision making accuracy parameters would be a better way to do it. My suspicion would be that value based outcomes would be more spread and therefore provide more of the extreme results at both top and bottom. I suspect the G&R approach would be more concentrated. I would expect the conclusion would be that you give yourself a better chance at a higher score (even a winning score) by using a G&R approach.

 

[warewolves]

Thanks for taking an interest.

Yes it has been a fair challenge to get to this point however I have been able to build my knowledge over the last few years and feel as though I am mostly on top of it now. The biggest barrier now is the software I have access to. Generally my ambition now tests the limits of that software and getting higher powered software would come at a fair cost both in terms of dollars and another learning curve.

Anyway back to your question. My own personal view is that whilst I find this an interesting academic exercise I don't think it offers too many strategic insights into how to succeed at this game. Generally what you see is that it trades players that are about to have a lull (either form, role or injury) for players that are about to go on a strong run. In my view that is impossible to predict and anybody attempting to do so will fail.

The problem with the value based approach is that relies on you getting the vast majority of value selections correct and history tells us that is nigh on impossible. If your success rate isn't high enough then you consign yourself to a miserable season because you will slip off the pace and there aren't enough trades available to correct the situation, be it switch to G&R or different value selections. I don't really think that the max score scenario supports that a value based approach can work because it so extreme and therefore unrealistic; however equally I don't necessarily think it means it can't work either.

In order to truly assess the merits of each approach I think some type of probabilistic modelling approach which builds in realistic decision making accuracy parameters would be a better way to do it. My suspicion would be that value based outcomes would be more spread and therefore provide more of the extreme results at both top and bottom. I suspect the G&R approach would be more concentrated. I would expect the conclusion would be that you give yourself a better chance at a higher score (even a winning score) by using a G&R approach.

Many thanks @Beg2Differ, you have made some very good points there.
My career has been in programming/IT and I have never had to try to code anything remotely as difficult as your SC optimiser – must have been very satisfying to get it all to work.

Maybe I’m just too stubborn for my own good, but I’m clinging to the idea that your findings must provide some insights into the best way to set up our starting squads. I’ve copied in your optimal starting line-up again. Dots against players I had in my 2021 squad after Rnd1.
I agree we could never get anywhere near the optimal solution and also agree that it “finds” players who are about to go on a hot-streak – sometimes even a singleton high-score. But if I look at the players who are traded in, it’s a virtual roll-call of the 2021 SC heroes.
DEF – Stewart, Laird, Dawson, Crisp
MID – Lyons, Miller, McCluggage, Parish, Oliver, Steele, Petracca, Wines, Mitchell, Walsh
FWD – Danger, Hall, Zorko
RUC – Gawn, Marshall, Nicnat, Darcy

Now I suppose, this is expected, but you must then ask, how did we generate the value to trade-in all of these high-priced guns – guns that weren’t in (couldn’t fit in) the original squad. From my observations most of it came from the MP brigade & not too much of it came from the rookies. In fact, there were few if any rookies that were traded-in – so no reliance on trading up thru a progression of 2-3 rookie players. If you start a lot of rookies, there just aren’t enough trades to bring in all those guns.

Now that just might emphasise the point that the distribution of outcomes will likely find that MP squads at the extremes and G&R squads clustered in the middle (upper-middle?). Trying to add a layer of probabilistic variables would be a very big ask.
Perhaps when (AND IF) you are setting up your model for 2022, you could run it in 2 modes – IF it wouldn’t be too much extra work @Beg2Differ! One as per now and another run thru whereby the starting squad was forced into a typical G&R structure – say 14 above $500K, 14 at or below $200K and 2 others. It would be interesting to see how close the best G&R team could get.

 

[Beg2Differ]

Many thanks @Beg2Differ, you have made some very good points there.
My career has been in programming/IT and I have never had to try to code anything remotely as difficult as your SC optimiser – must have been very satisfying to get it all to work.

Maybe I’m just too stubborn for my own good, but I’m clinging to the idea that your findings must provide some insights into the best way to set up our starting squads. I’ve copied in your optimal starting line-up again. Dots against players I had in my 2021 squad after Rnd1.
I agree we could never get anywhere near the optimal solution and also agree that it “finds” players who are about to go on a hot-streak – sometimes even a singleton high-score. But if I look at the players who are traded in, it’s a virtual roll-call of the 2021 SC heroes.
DEF – Stewart, Laird, Dawson, Crisp
MID – Lyons, Miller, McCluggage, Parish, Oliver, Steele, Petracca, Wines, Mitchell, Walsh
FWD – Danger, Hall, Zorko
RUC – Gawn, Marshall, Nicnat, Darcy

Now I suppose, this is expected, but you must then ask, how did we generate the value to trade-in all of these high-priced guns – guns that weren’t in (couldn’t fit in) the original squad. From my observations most of it came from the MP brigade & not too much of it came from the rookies. In fact, there were few if any rookies that were traded-in – so no reliance on trading up thru a progression of 2-3 rookie players. If you start a lot of rookies, there just aren’t enough trades to bring in all those guns.

Now that just might emphasise the point that the distribution of outcomes will likely find that MP squads at the extremes and G&R squads clustered in the middle (upper-middle?). Trying to add a layer of probabilistic variables would be a very big ask.
Perhaps when (AND IF) you are setting up your model for 2022, you could run it in 2 modes – IF it wouldn’t be too much extra work @Beg2Differ! One as per now and another run thru whereby the starting squad was forced into a typical G&R structure – say 14 above $500K, 14 at or below $200K and 2 others. It would be interesting to see how close the best G&R team could get.

This is just a hobby for me even though I do have a bit of an aptitude for it. The SC optimiser is actually a bit of a bonus because my initial objective in learning about this stuff was in regard to another of my hobbies. I have been heavily involved in community sport (basketball) administration for many years and wanted to find a way of automating fixturing which is quite complex. That passion is something I shared with my father which started in the 1980s (I was in my teens) when we got our first PC. I have largely achieved that now nearly 40 years later and that has been immensely satisfying. Unfortunately my dad passed in 2015.

You do make a good point in regard to cash gen coming from value picks and the optimal team has virtually no downgrading to rookies. Having said that I think the starting team more closely resembles a typical starting team than it has in other years.

I have considered trying to come up with an alternative scenario along the lines of the one you have suggested but ultimately it becomes very subjective about how you restrict the trading, to the point of being contrived. The other issue is that it already takes 70 odd hours to solve a full season and adding further constraints only means it will take longer. If it solved quicker you could play around with this type of scenario but with it taking so long to see results it becomes impracticable.

Having said all that I could certainly set it up for the early rounds of 2022. I will give it some thought.

 

[warewolves]

This is just a hobby for me even though I do have a bit of an aptitude for it. The SC optimiser is actually a bit of a bonus because my initial objective in learning about this stuff was in regard to another of my hobbies. I have been heavily involved in community sport (basketball) administration for many years and wanted to find a way of automating fixturing which is quite complex. That passion is something I shared with my father which started in the 1980s (I was in my teens) when we got our first PC. I have largely achieved that now nearly 40 years later and that has been immensely satisfying. Unfortunately my dad passed in 2015.

You do make a good point in regard to cash gen coming from value picks and the optimal team has virtually no downgrading to rookies. Having said that I think the starting team more closely resembles a typical starting team than it has in other years.

I have considered trying to come up with an alternative scenario along the lines of the one you have suggested but ultimately it becomes very subjective about how you restrict the trading, to the point of being contrived. The other issue is that it already takes 70 odd hours to solve a full season and adding further constraints only means it will take longer. If it solved quicker you could play around with this type of scenario but with it taking so long to see results it becomes impracticable.

Having said all that I could certainly set it up for the early rounds of 2022. I will give it some thought.

If you do manage tweak something for next season, I'll be very interested to see what the optimiser throws up.
It did occur to me that it might get tricky to define a G&R team to everyone's satisfaction as there is likely to be some blurring around the edges of any definition & I am no authority on G&R set-ups! Am going to run with as many value options as I can find in 2022.

To me this is a really interesting & important SCS forum/thread, so again thanks for putting in all the time & all the effort - much appreciated.

 

[Beg2Differ]

Maximum score for the first 6 rounds is 8898 (avg 1483) which compares to the current leader who has 6078 (avg 1013).

Note this in an updated posted. I realised I made an error in the first run.