Using AI To Build Better Items ft.

[@BrokenByConcept]

Thanks to Xpetu, use code "BBC2024" to get 50% off your first month off Coachless (valid until the end of 2024). coachless.gg

[@rokasb9441]

Hi, i think it would be a cool idea, (both in itself and to promote podcast) to take some interesting clips from it, about some talking points and upload them as youtube shorts ;)

[@xPetu]

An absolute pleasure visiting the pod again! Thanks for having me.

[@Gormezzz]

Oie oie oie

[@scalpen]

@xPetu what do you think about the "Combined Sets" tab on lolalytics? Instead of referencing a naive item winrate, it takes into account probabilities for games where you were unable to finish your item and games where you were able to build more items past the items listed. It gives you a kind of "true" winrate disregarding sample size. Is that something you are taking into account currently? i.e. if I build Rabadons third item on a mage, its very rare that I complete the item before the game ends, whereas if I build a cheaper item like Cryptbloom, I am much more likely to complete it. So, if you only do analysis on the item once it is completed, your win-probability would be disproportionately affected (in this example it might be the case that buying Rabadon's is projected to have a higher "win probability" because if you are actually able to complete the item or afford to buy its components, the end result is a stronger item, but in games where you can't complete the item you would have a much lower win probability than buying Cryptbloom).

[@Salaryman_]

I wonder if “mathematically correct builds” align with higher added win probability builds? Hmmm 🤔

[@benamende5897]

Would be really interesting to see the data on mejais when purchased from different numbers of dark steal stacks

[@MiniDonbeE]

Petu... I was checking shyvana on coachless and saw that the first two winningest items are ad, but the last 2 winningest items are ap.... I assume this means that if you go AD shyvana you are likely to end the game early and if you go ap you are likely to end the game later on, OR does it mean I should be playing ad shyvana early and transition into a half bruiser half caster shyvana later on? Im REALLY curious if the right answer is that, kraken into sundering into, maajais into horizon REALLY is THE winningest build. Because if it is then that changes a LOT of things in my mind.

[@Norrman_RT56]

Last time I was this early, Blue Ezreal was the Meta build.

[@Pyrrha_Nikos]

Takeaway from the episode: build mejais when losing, Trailblazer is op and feelings are overrated. Petu is the goat

[@FiveDollaPhillip]

The way I understand the mejai’s from behind debate is, in order for you to claw your way back into a game that you are extremely far behind in is to get picks and hope that stalls enemy tempo enough. If you keep getting picks mejais also does well. It’s like a playstyle synergy between what is required to win from behind and what is a good situation for mejais

[@gonzalotoscano3310]

Also you can understand it as a gamble, why have you $10 dollars and use it for something that costs $10 when you can gamble and turn those $10 into $100. With mejais from behind you are buying the possibility of getting more gold than you should from your item.

[@Lkabss]

"Better to roll the dice than resign to fate" - Me, just a second ago.

[@CollectorVX]

I think LS is absolutely correct here tho. If only one pro has success with a certain item or build, makes a guide or streams about it and the masses start buying it, the stats for it WILL go down almost always, in fact this is basically how every single meta strat gets known. Of course this doesn't affect the actual effectiveness of that item whatsoever, especially not for those primary skilled players. But it's worth pointing out that coachless does NOT use those extremely skilled players as basis for the item analysis, simply because the sample sizes for high elo in major regions are just too small to allow this, and as a result of that DOES get affected by this 'popularity' effect. This is a major gripe I have with coachless in general - if the majority of data comes from low- to mid elo, far away from where you will actually get challenged on how optimal your items are, there may be a lot of significance to the results, but there might also be none at all, and there's no way to know which is the case. In other words, in my eyes the builds of a silver player is completely worthless data. You might say that statistically, surely the sheer sample size accounts for this and the most optimal builds swim to the surface. And that might be the case, but it might also NOT be the case - for example if - hypothetically- nobody in plat or below, or even for the sake of argument no human player on the planet would be able to use zhonias correctly, it would be deemed a dogshit item, while a deep learned machine might deem it best in slot. I don't know how coachless is supposed to be able to make meta-breaking predictions like these and others mentioned in the video, when there is - and this is important- NOT ALREADY PRECEDENT for that build IN THE SAMPLE, so (i assume) at best plat or emerald. I'd love to know more about the briefly mentioned anti-biasing tech employed here and how coachless manages to discern between correlation and cause, and how you arrive at the momentary win% stat in the first place, if not purely based on the same hindsight- stats lolalytics & co uses. Every time I hear about coachless I wonder this and it's never really answered- HOW can you decide whether you win because your item is good, or the item has good wr because you happen to buy it while you were winning anyways. I have no clue about statistics, so if that is truly possible, this tool is huge. If it's not, this is the most elaborate results-oriented stat site in history...

[@xPetu]

Hi! Thanks for the interesting comment, here's a point-by-point rundown of your concerns: > coachless does NOT use those extremely skilled players as basis for the item analysis We have a tier selection system so you can choose to only look at e.g. Diamond and Master+ games (or any other subset from Gold to Master+). It's true that the sample sizes are lower in high elo, but for popular champions you find plenty of games in just Diamond and Master+. We also have a graph that shows you the changes in Win Probability Added (WPA) with respect to rank, allowing you to see the changes in "high-skill" items like Zhonyas, Deadman's Plate, Mikaels etc. > I don't know how coachless is supposed to be able to make meta-breaking predictions like these and others mentioned in the video The product/method alone is (currently) not able to make "meta-breaking" predictions, this is where you need to combine your game knowledge and theory (as discussed in the podcast) with the data. However, Coachless does allow us to find general trends in item balance, which allows you to find "overpowered" items MUCH MUCH more efficiently than the existing methods. > I'd love to know more about the briefly mentioned anti-biasing tech employed here and how coachless manages to discern between correlation and cause Firstly, the WPA is already a debiased statistic, because we remove the effects of the purchase situation from the items strength indicator. The problem with win rates is that win rates include both the purchase situation and the effect of the item. Whenever we give an estimate of item strength (=WPA) with some situational filters, e.g. "Aatrox vs Magic Damage", we look at the general statistics of Aatrox and Aatrox vs Magic damage specifically and remove those effects to isolate the item's effect in that particular scenario. This is done by "centering" (mathematical term, means subtracting the general mean) the data in that specific situation. > how you arrive at the momentary win% stat in the first place We developed a deep learning win probability model that is trained on millions of games. The model takes as input the game state and outputs the win probability. The accuracy (% of correct winners predicted) ranges from roughly 60% (early game) to 80% (late game) and the expected calibration error (how far is the win probability from the "true" underlying win probability, i.e. proportion of games won from similar states) is around 1%, meaning that the model is very reliable but not perfect. Hope this clarified some things! let me know if there's anything else that you are wondering about!

[@CollectorVX]

@@xPetu Hey xPetu, thank you for taking the time to reply in detail. This is very insightful and clears a lot of things up, I was worried coachless was more or less - no offense- a lolalytics clone with better item navigation for the user, at least as far as the itembuilds part of it goes. I'm definitely seeing its potential, you guys have rly thought of everything! I'll follow coachless' career with great interest, amazing stuff

[@PlanADidntWorkOut]

when Miss Fortune Support was picked at Worlds, botlane became absolute terrorism

[@LiferSmite]

I love you three working together, only if the league community was made of this caliber of intelligence and conversation skills

[@triez7454]

Damn now I'm building mejai every game

[@allenzhu1347]

coachless seems pretty impractical still in making usable data. I think his take does account for a common source is bias in data which is win rate at moment purchased, but doesn't account for overbuying or underbuying of items. some items are only bought when they are really really good (not always dependent on winning and losing) like locket on enchanters therefore inflating their winrates, and some are possibly overbought like locket on tanks supports which deflates their winrates. common sense could tell you locket is better for tanks than enchanters but stats say otherwise because they are bought in different situations (independent from win probability)

[@goldythefish36]

It's still a work in progress but I think even if it has its faults it is still giving more information behind the data and statistics. Compare your example with existing sites. They would spit out similar information but with even less reasoning. He talked about adding new features with components, enemy comps etc. So I agree items still can be inflated somewhat, but with more useful information you may be able to draw better conclusions faster instead of just copy and pasting builds. Using sample size, win at purchase and win increase could be very valuable together. A low sample size high win rate increasing item could be bought specifically to counter an enemy comp. And then the starting win rate may follow the average for the champ. Then applying logic to it may lead you to a new understanding which would be much harder without the win prob change.

[@Lkabss]

Wouldn't be surprised if he knows and just can't use do much. My last project for my bachelors was on this, and I got nowhere. Had to dumb the whole thing down orders of magnitude to finish. Working with league was much more complicated than I thought.

[@aznir2427]

I prefer the darkmane ap garen build but to each their own I guess

[@Hylofear]

6. Get 10 kills as promised. 7. Lose anyway.

[@WB650]

Thanks for getting Viktor from arcane on the podcast!

[@Jadenity]

Thank you so much for this gift. As a software architect interested in machine learning, this episode was fantastic. I could listen to this all day long. I'd love to see more content around this subject!

[@thebordercollier]

Thanks so much for doing this video! I've been a Petu fan for years (the OG off-meta item builds / mathematical analysis of "correct" builds was goated content). @xPetu, @BrokenByConcept, and @Azzapp are my Holy Trinity for the mental and social side of the game. League will be a better place if more players adopt your collective value systems. Y'all are wonderful and this interview gave me so many answers to things I've been wondering about what my favorite League Player has been up to since graduation. Thanks again!

[@AxeDropper4000]

Curtis and Nathan just nodding along trying to follow the thesis like the rest of us.

[@lachiemunro7]

Follow the data! The DMP Akali explaination as well as the DETAILS about the gameplay patterns of akali and dead mans make this a great segment. So interesting.

[@confusednerd3171]

This episode is gonna go down in history for revolutionizing item builds in league of legends!!

[@MIIC2]

BenTbeyondrepair just published a video In Defence of Toxicity. Where he argues that non toxic players are not competitive enough to understand. Thanks for Broken By Concept I know better. If you are truly competitive you respect the need to stay calm and take breaks so you can perform at your peak to climb. Thanks for another great video! :)

[@VigilTheProtogen]

Thumbnail looks like an Ultimate Bravery loadout ngl. I am studying AI for my master's degree right now, and I have some trepidations about these "math correct" builds because the client may not understand how to play around the unfamiliar itemization. Nonetheless, I would love to see a product like Coachless evaluating pro player builds while matches are going on, and hearing whether commentators favor the human or the AI build

[@Jadenity]

I'm surprised the API doesn't provide minion location information. I wonder how much more accurate the model could be if Riot also provided minion information. Minion wave states and jungle camp timers have a huge impact on decisions people make, and I suspect a deep learning model would pick up on that.

[@juliengauthier523]

Xpetu IS A LEGEND OF youtube for real ! I dont watch so much otp content but xpetu is just insanly good to any form of explanation. Love this man. Big love from france

[@skipp3252]

I wonder how much item purchases like mejais influence the win probability through mindset of the player unrelated to the item stats. Like a player that buys mejais might play out situations very differently than the same player with a different item. And not only with mejais because you dont want to die. Generally if you build more glass cannon you may start playing more scared or cheesy. If you get deadmans plate on Akali you might put more emphasis on your movement within your combos than if you didnt have it. These things may even be greater difference makers than the different stats.

[@elliotbennett3917]

As a support main, there's a big difference between the general "skill" of Enchanter players, Mage players, and Engage players. The role is exactly the same but your job in the comp is vastly different because you are expected to be utility/dmg/tank based on your champion. This is why it's commonly known that enchanter mains are "inflated" in elo and struggle to perform mechanically, roam correctly, ward effectively, because the threat of death is much lower and you are rewar5ded for playing very very defensively. When I play against enchanter players it's usually a very easy lane/game, because I main engagers, and the synergy of me playing better as well as enchanters being counters by most engagers

[@haroentmimi]

The BBC podcast makes every monday better

[@Return989]

I think Mejas is a very specific case. Its winrate from behind is probably so crazy, because instead of slowing bleeding out you basically, all in on winning one teamfight and then the swing in the game is higher, because additionally to the shutdowns you also get alot of "free" AP. But that probably is also only true because of fairly low samplesize right now.

[@gantonstepfire3716]

Love this, love both of these channels. Great Talk! 👍

[@Jadenity]

Regarding the "sacred League" discussion: I would LOVE to have an AI coach to tell me the best thing to do at critical points in the game. I feel this would help me improve drastically. However, I think this is similar to cheating in chess. Even having a chess engine give a grandmaster an indication that IT IS a critical point in the game is considered cheating, not to mention getting the best move from it. So, while I think these tools would be incredibly useful, I don't think it would be widely accepted in the community.

[@raiohkyros]

love xpetu and bbc collabs!

[@Salaryman_]

xPetu was a sick guest, more like this and more xPetu please 🙏

[@RocoPwnage]

Normally I'm like you guys and I don't give a shit about items at all but this was still a very interesting listen. I'll definitely be more on the lookout now for items that don't "feel" good but are actually broken.

[@dannytamale860]

Just started playing moba for the first time in my life. I am trying to consume all the information to be able to learn league. XPETU MY GOAT. Love the content you guys ❤️

[@Xuor13]

So cool to hear xpetu is working on componant too because I felt like looking at the win probaility added without looking at why some item have better win probability bf don't tell the whole story. Some item like mejai have high winprob bf because people buy it when ahead but some just have really good componant pathing like maybe timat for reksai we saw early in the video

[@pseudophi]

One of the first time I hit Diamond on Rek I was running Ingenious Hunter Guardian Angel rush with HoB. Time to get back to the lab with cosmic insight GA rush XD

[@eddierosario7811]

The great thing about league is that its getting so old now that the OG players are now so invested that the experience level is maturing from mindless gameplay to serious analysis. Most of us have over 5k hours so we can say we have "degrees" in this game

[@thewordisbranch]

The robots are winning

[@AmericanRatBoy]

Petu my goat

[@Kohlroulade]

mejas is the number one item for people who know they are better than the other players in the game. some people naturally have this confidence BUT Smurfs definitely have this confidence! Smurfs fake the data! Only because he buys mejas from behind SilverLarry123 wont become a super confident highelo smur. Maschine learning only works as long as humans can use what the machine is delivering. GA is the highest winrate adc item because u buy it when u got enough dmg already. its not the reason u win! u buy it because u already win!

[@byeguyssry]

A question I have for xPetu is, how can you verify if Coachless works? I'd imagine that the win rate of people who buy an item due to Coachless's recommendation is not necessarily the same as the win rate of people who buy an item without looking at Coachless (for instance, people who use Coachless might spend more time improving at League compared to other players; alternatively, people who use Coachless might be too focused on the recommendations in a vacuum that they fail to take into account the game state). Therefore, if Coachless is flawed, how can you detect it?

[@philiphession8700]

Best episode ever (for me anyways!). Period. Thank you for doing this.

[@madaraIIII]

Wait Petu was on bbc before??? Amazing Podcast btw, it's enhancing my inferior neural Network alot. I want to join the program soon.

[@death7731]

Regarding all relevant topics, we should incorporate visual representations of the matrix in the next one :)

[@TonyYang-v9x]

The link for coachless is not in the description :c

[@Masterpouya]

Is there a place where we can read xpetu works on how the algorithm works and such?

[@moresah]

The most important thing is to understand how you should play with your build, not to always find the best build.

[@ardiansyahkesuma7464]

Welcome to dojo

[@jackweatherhead8171]

Chess

[@saminvalidbaka]

Xpetu kinda lucky to be in the #228 episode

[@ExclusiveExcellence]

44:40 nah man there has to be a statistical bias or something where GA is normally only bought last after the ADC already has full build so it appears to be responsible for the win percentage when in reality the given player would win regardless of GA, but idk like they said I guess the stats don’t lie

[@Matkatamiba]

GA is sometimes a perfect item when their wincon is a super over-extending dive to kill the adc. BT usually performs better, since GA can mean just dying immediately on respawn again if you buy it in the wrong situation.

[@ExclusiveExcellence]

I will say I made this comment before he showed coachless on screen

[@HebiSnake]

It could also be that buying it earlier will give it more time for the passive to inpact the game, who knows right?

[@josecajigas9210]

So this is why my 0/12 lux builded mejai’s…

[@Salaryman_]

Surely the meta an AI would go to is 2 mid laners for better objective control and being less susceptible to jg ganks.

[@PlanADidntWorkOut]

- 5 man mid with perfect tower dives - 5 man roaming deathball with perfect fog of war + jungler tracking - 5 man level 1 triple proxy - 5 river shens

[@Salaryman_]

@ I think you have some interesting ideas, but I don’t know how successful they would be at elite level

[@PlanADidntWorkOut]

@@Salaryman_ hey who can say? 5 ais with superhuman micro, synergy, information processing and decisionmaking could find something totally off-the-rails like the dota ai did. I mean we've seen crazy strats like double smite on top/mid/support, funnel, ady's janna top and inting sion dominate high elo - League is far from a solved game

[@AzizTheGazele552]

Came to most recent video hoping to understand something Coach Curtis talks about a lot, what do you mean by "lol states"?

[@caruja5301]

@AzizTheGazele552 He's actually saying 'lull states' - meaning there is a slow down or pause in the action where you aren't occupied by anything else, and have spare time to think about the state of the game as a whole (maybe because you died, maybe because you based and are walking back, maybe you cleared your current wave and have nothing else to use your brain on right now, etc)

[@commenthandler]

dark seal is soo good on defensive ad champs w high ap ratios

[@commenthandler]

for buying it on xin in the right games makes you insanely strong

[@commenthandler]

also %based things completely break itemization options

[@nosc_c]

Challenger OTP is just an inflated Master player

[@Alend222]

The builds work I tried the Deadman's Akali build and it's really broken

[@Carlisert]

Yesterday I listened the last xpetu episode, what are the odds

[@nessbrawlaaja]

torilla tavataan

[@innemannen]

I tried buying the highest "win probability added" items for Orianna and it was absolute dogshit. Idk if I am just misunderstanding the data or what but rod of ages cosmic drive orianna is terrible.

[@ForzaOwnz]

Give the "Power of evil" build a go, roa into nashors

[@odawgreallyinthefield9179]

awesome episode! xpetu definitely getting a sub from me

[@mavriksc]

So riot just added a mechanic around FB tower and objectives and there's betting on those things 🤔

[@Thedoctorr5]

I would find it sad if the meta changed because people copy and AI strat instead of meta changes coming from human discovery.

[@garagavia]

Jungle timer should be built in to the game,! Wtf is Riot doing to ignore this?? Tilts me to have to use a 3rd party app for that

[@Junior305able]

Please do attack speed sion

[@xanix6862]

GIVE ME QIYANA DATA!!!!!

[@EVERYTYME]

Ive built majeis on galio every single game since Xpetu released a YT short on it a month ago. That item is not ok! 100% should be removed from the game

[@PkerHaezel]

Great video, Petus work seems super interesting. Think you forgot to put the link in the description though =)

[@tatybara]

ai zzz

[@rxpe]

ai :tf:

[@AlphadystTFT]

Two coaches not in the same physical space, and Petu without cam. This makes the podcast completely unwatchable.

[@saki1333]

The problem I have with these models is that even if on average d cap would be best for example, Vs 5 tanks void is still the way. they're nice to have another perspective but not something I'd go by.