AI to Improve computer GM

[Deft]

Has there been any work on using AI to improve computer GM performance? I imagine running simulations over and over using neural networks learning how best to organize lineups, promote and demote minor leaguers, as well as trading, could dramatically improve computer GMs.

[Bobfather]

Doesn't AI need to run in the cloud?

[Haiku]

The game engine which is programmed to simulate and make decisions on behalf of both the player and the opponent, is called AI. It is an integral part of the game's code. I can't see how an external AI could be used to improve the in-game AI and change the code.

[Syd Thrift]

People keep saying this, and the bigger issue to me is, I don’t know why you’d want that. An AI type setup is liable to try anything - batting the worst hitter on your team leadoff, playing guys out of position, running out starters for 3 innings, etc. - and then will re-evaluate based on either direct input, which means lots of people having to tell the AI that it messed up, or you plant a thing in where it “knows” it did well based on wins and losses.

That’s not a method to create the Ultimate Baseball Manager, that’s a method to create the ultimate exploiter of loopholes in the game itself.

[MathBandit]

Originally Posted by Syd Thrift

People keep saying this, and the bigger issue to me is, I don’t know why you’d want that. An AI type setup is liable to try anything - batting the worst hitter on your team leadoff, playing guys out of position, running out starters for 3 innings, etc. - and then will re-evaluate based on either direct input, which means lots of people having to tell the AI that it messed up, or you plant a thing in where it “knows” it did well based on wins and losses.

That’s not a method to create the Ultimate Baseball Manager, that’s a method to create the ultimate exploiter of loopholes in the game itself.

(and also if they did somehow manage to create the Ultimate Baseball Manager, 95% of the fanbase would complain that it is both very unrealistic- since it won't look anything like most MLB lineups/rotations/etc- and very unfair- since it will significantly outperform user lineups who think doing what MLB teams is doing (great hitter batting 3rd, 5 starters all expected to throw 5+ innings, etc).

[Deft]

Originally Posted by Syd Thrift

People keep saying this, and the bigger issue to me is, I don’t know why you’d want that. An AI type setup is liable to try anything - batting the worst hitter on your team leadoff, playing guys out of position, running out starters for 3 innings, etc. - and then will re-evaluate based on either direct input, which means lots of people having to tell the AI that it messed up, or you plant a thing in where it “knows” it did well based on wins and losses.

That’s not a method to create the Ultimate Baseball Manager, that’s a method to create the ultimate exploiter of loopholes in the game itself.

My assumption was not to have it be an active AI running from the cloud and exploiting all of the game loopholes in the customers hands but use it during development as an AI tester to exploit hidden loopholes to close before users find them, and improved post-launch AI algorithms through pre-launch testing.

[Syd Thrift]

Originally Posted by MathBandit

(and also if they did somehow manage to create the Ultimate Baseball Manager, 95% of the fanbase would complain that it is both very unrealistic- since it won't look anything like most MLB lineups/rotations/etc- and very unfair- since it will significantly outperform user lineups who think doing what MLB teams is doing (great hitter batting 3rd, 5 starters all expected to throw 5+ innings, etc).

Yeah this is also true and I really should emphasize a particular point here that makes AI differ greatly from human managers: all real-life managers, even the innovative ones, play to not get fired, not to experiment and run the most optimal setups. We've seen a lot more use of openers and some reworking of lineups in recent years because the data back up the change but also because teams have been successful with them. This is really important, I think: in the late 80s we had this big change of opinion on the role of bullpens, not because that was "optimal baseball" but because the A's had 3 short relievers who carried them in the regular season and the playoffs. That's one of many examples.

If, on the other hand, a human manager decided to run a 3-man rotation and give all of their starters 80 pitches, if they fail to win - even if the failing to win has nothing to do with this strategy - that manager would get fired and would find it very hard to find a new job. Even if they won but one of their pitchers got hurt, there's a good chance they'd get fired for it (unless they won a lot) because all anyone would talk about is how their weird strategy is "ruining the game" or what have you. It's only natural that human managers play conservative: taking big risks means you have to find work outside of baseball.

Baseball wisdom is built on like 150 years of assumptions and results. Very often, stuff that we think is new thinking is in fact just repurposes old thinking - one big example for me is the idea behind FIP, where it was an accepted thing for decades that the job of a pitcher is to let their fielders make plays and not get themselves into trouble, but I don't know, a couple generations of guys growing up on Strat-o-Matic or something made some statty types think that pitchers have an ability to affect hits on balls in play. Baseball Prospectus came out with a book around a decade or so ago that talked about how a lot of stat guys were coming around to accepting that conventional wisdom had a lot of good data behind it in many cases. But bucking that wisdom, in baseball or in any other sport, comes with a cost.

[Ktulu]

Originally Posted by Deft

My assumption was not to have it be an active AI running from the cloud and exploiting all of the game loopholes in the customers hands but use it during development as an AI tester to exploit hidden loopholes to close before users find them, and improved post-launch AI algorithms through pre-launch testing.

It's easy to use a bunch of AI buzzwords, but in a practical sense how would you actually do that?

[Deft]

Originally Posted by Ktulu

It's easy to use a bunch of AI buzzwords, but in a practical sense how would you actually do that?

Create a standard test league and interface for commands from your OOTP program and the neural network with variables that can be modified (example- trading of major league players) and defined outcomes (ex. below budget, and wins). Then have the neural network conduct millions of sims of that year to optimize the outcomes. When done, evaluate the trade decisions for trends, (ex. trading timing, player type traded, # of trades) to see if their might be obvious loopholes the neural network is taking advantage of or whether the activity is as expected from a regular human user.

[LacSlyer]

This simply isn't how AI works right now. You can use it to help you code snippets, but it isn't ideal for using it to design a program as intricate as OOTP, let alone any other video game.


People need to also understand that it will be a very long time until AI in video games is capable of competing with a human because of how games work. This isn't like chess or Starcraft, where there are limited moves the computer can execute. In a game as advanced as OOTP you can't have an AI that looks at what they want to do right now and consider how that'll effect what they do in an hour that seems completely unrelated. All strategy games have this issue and will continue to have this issue for a while because it's really not that simple to fix.

[Ktulu]

Originally Posted by Deft

Create a standard test league and interface for commands from your OOTP program and the neural network with variables that can be modified (example- trading of major league players) and defined outcomes (ex. below budget, and wins). Then have the neural network conduct millions of sims of that year to optimize the outcomes. When done, evaluate the trade decisions for trends, (ex. trading timing, player type traded, # of trades) to see if their might be obvious loopholes the neural network is taking advantage of or whether the activity is as expected from a regular human user.

I mean the first half of the first sentence is probably months of work just to setup... and then you're assuming a magic neural network can actually generate useful data that doesn't take many more months to analyze and that it actually tells you something that you don't already know.

After all that you're still at the hardest part... programming the AI to be better.

[jksander]

Originally Posted by Syd Thrift

That’s not a method to create the Ultimate Baseball Manager, that’s a method to create the ultimate exploiter of loopholes in the game itself.

lol ... that summed it up pretty well for me

[uruguru]

Originally Posted by Syd Thrift

That’s not a method to create the Ultimate Baseball Manager, that’s a method to create the ultimate exploiter of loopholes in the game itself.

That's exactly right. If there is an exploit to win your games, the AI will eventually find it. Then players will complain that the AI found a way to lock up all of the catchers in the league or something and so the player has to run a LF in the catcher slot every night.Or something.

So then the devs have to fix the exploit that the AI is abusing, and then spent tons of money for server time to regenerate the AI on the updated code, only to have it find the next exploit.

A game as complex as OOTP could literally take many years to find and close every possible exploit. It's not a viable business model.

[Matt Arnold]

Balancing "play to win" vs "play baseball" is definitely always a challenge... How do we make a manager do dumb decisions, but the *right* dumb decisions?

[TheChief]

Technically anyone could sign up and teach one of the open Ai's on how to be a real life MLB manager.

You would have to teach it, if it' not being done already. Something like load up all the lineups, results, roster moves, score cards, etc into the Ai. It would then learn how to be like a real life manager.

It wouldnt be depended on OOTP or a game but real life info. It probably would never be able to be a commercial product or anything like that.

Steam also has it against it's ToS rules to use AI artwork and Ai Audio. Because there hasnt been any laws yet about this area. Remember for the neural network to learn any skill you are teaching it based upon a real life persons work, whether it is their voice or art style or even writing.

A lot of the hollywood strike was over Ai. Especially in the so called reality tv programs. As they still work on scripts. And since they all follow a pretty specific format it is actually an area an Ai writing would excel at. Of course you would need an editor but you just eliminated a lot of human and pay. So Hollywood was wise to to make this an issue.

Ai has gotten really good at Artwork as well. It has been is kids novels and other books that sell many copies. Wizards of the Coast the company who owns D&D had Ai artwork in some of its guides or whatever your call them. People made a stink so I believe they were replaced.

A lot of spam calls are actually Ai voice chatbots now and its pretty hard to tell. You can google videos. Some spam call will be about whatever topic voting and if you ask the "person" to give you a summary on some random topic like how to make a chocolate cake it will just go right into it.

It would be cool if some huge baseball fan did it as an experiment just to see what it would be like comparing to real managers. Just for fun and entertainment purposes

[uruguru]

Originally Posted by Matt Arnold

Balancing "play to win" vs "play baseball" is definitely always a challenge... How do we make a manager do dumb decisions, but the *right* dumb decisions?

Dude, I can so relate to this. There are two schools of thought to AI development... 1) it's a game and the AI should play to beat you. 2) it's a simulation and the AI should play to immerse you. I wanted #2 but the AI developer wanted #1.

We had big arguments about this with rotp and I eventually ended up with multiple selectable AI's.

[iridescence93]

Originally Posted by Syd Thrift

Yeah this is also true and I really should emphasize a particular point here that makes AI differ greatly from human managers: all real-life managers, even the innovative ones, play to not get fired, not to experiment and run the most optimal setups. We've seen a lot more use of openers and some reworking of lineups in recent years because the data back up the change but also because teams have been successful with them. This is really important, I think: in the late 80s we had this big change of opinion on the role of bullpens, not because that was "optimal baseball" but because the A's had 3 short relievers who carried them in the regular season and the playoffs. That's one of many examples.

If, on the other hand, a human manager decided to run a 3-man rotation and give all of their starters 80 pitches, if they fail to win - even if the failing to win has nothing to do with this strategy - that manager would get fired and would find it very hard to find a new job. Even if they won but one of their pitchers got hurt, there's a good chance they'd get fired for it (unless they won a lot) because all anyone would talk about is how their weird strategy is "ruining the game" or what have you. It's only natural that human managers play conservative: taking big risks means you have to find work outside of baseball.

Baseball wisdom is built on like 150 years of assumptions and results. Very often, stuff that we think is new thinking is in fact just repurposes old thinking - one big example for me is the idea behind FIP, where it was an accepted thing for decades that the job of a pitcher is to let their fielders make plays and not get themselves into trouble, but I don't know, a couple generations of guys growing up on Strat-o-Matic or something made some statty types think that pitchers have an ability to affect hits on balls in play. Baseball Prospectus came out with a book around a decade or so ago that talked about how a lot of stat guys were coming around to accepting that conventional wisdom had a lot of good data behind it in many cases. But bucking that wisdom, in baseball or in any other sport, comes with a cost.

c
cool thing about OOTP though is you can try these things without worrying about losing a RL job or being laughed at. It would be pretty cool if the AI got to a point where it could also come up with weird out of the box strategies some of which may outperform conventional baseball.

[highandoutside]

Originally Posted by Matt Arnold

Balancing "play to win" vs "play baseball" is definitely always a challenge... How do we make a manager do dumb decisions, but the *right* dumb decisions?

haha, a true dilemma!

Personally, I'd love to play the AlphaGo of OOTP. It'd be great for a fictional league and if it didn't follow "traditional" baseball intuitions, so be it. The challenge would be fascinating.

[Ataraxzy]

Originally Posted by Bobfather

Doesn't AI need to run in the cloud?

Not the kind of AI we're talking about here. This is standard machine learning stuff, which can be done, albeit slowly, on a laptop.

Originally Posted by Ktulu

It's easy to use a bunch of AI buzzwords, but in a practical sense how would you actually do that?

This type of game is actually brilliant for using an ML algorithm on. It's got TONS of stats to correlate, and a single, all-encompassing survival function: wins.

Been awhile since I did any ML stuff, and it was all genetics related, but, it's not terribly difficult, provided you can easily separate the game engine from the display engine (which, while a PITA, isn't impossible).

NOTE: I'm eliding a LOT of technical details here to give just a high level overview

Start it off with its basic in-game variables: batters, pitchers, their individual strategies & so on.

Get the ML working well at the scale of a single game, permuted a LOT of times.

Do some data analysis, see if you can identify any local maxima it might be stuck in.

Add more variables, until you have it managing everything about a game.

Do more data analysis.

Next, new ML: trading.

Same deal, use the basic management engine in the actual games to keep things equal.

LOTS of permutations and data analysis later...

Join the two and see what happens.

Might be a really fun project, honestly.

[Pledatchger]

Originally Posted by TheChief

Technically anyone could sign up and teach one of the open Ai's on how to be a real life MLB manager.

You would have to teach it, if it' not being done already. Something like load up all the lineups, results, roster moves, score cards, etc into the Ai. It would then learn how to be like a real life manager.

It wouldnt be depended on OOTP or a game but real life info. It probably would never be able to be a commercial product or anything like that.

Steam also has it against it's ToS rules to use AI artwork and Ai Audio. Because there hasnt been any laws yet about this area. Remember for the neural network to learn any skill you are teaching it based upon a real life persons work, whether it is their voice or art style or even writing.

A lot of the hollywood strike was over Ai. Especially in the so called reality tv programs. As they still work on scripts. And since they all follow a pretty specific format it is actually an area an Ai writing would excel at. Of course you would need an editor but you just eliminated a lot of human and pay. So Hollywood was wise to to make this an issue.

Ai has gotten really good at Artwork as well. It has been is kids novels and other books that sell many copies. In addition, he copes well with student assignments. A friend showed me essay writer helper and I was amazed how this chatbot gpt instantly gives a solution. I tested both mathematics and chemistry, excellent results everywhere. Wizards of the Coast the company who owns D&D had Ai artwork in some of its guides or whatever your call them. People made a stink so I believe they were replaced.

A lot of spam calls are actually Ai voice chatbots now and its pretty hard to tell. You can google videos. Some spam call will be about whatever topic voting and if you ask the "person" to give you a summary on some random topic like how to make a chocolate cake it will just go right into it.

It would be cool if some huge baseball fan did it as an experiment just to see what it would be like comparing to real managers. Just for fun and entertainment purposes

That would really be fun to see. But I think it would take a lot of time. I'm curious how successful it could be, I mean, how much of a competitor it could be to a real manager