AI question re trading

[Charlie Hough]

Originally Posted by CBeisbol

I'm not a programmer. But isn't the advantage of computers their ability to perform lots of calculations.

That's why computers are good at chess. They analyze lots of moves, and lots of moves ahead. Perhaps it's my lack of knowledge in the subject area, but it seems to me, the AI should (or could) excel at this.

Theoretically it's possible to code AI in a way that would allow it to plan ahead and think in sophisticated ways like a human GM can, but it would be staggeringly complex and would probably require far more coding and computational power than a chess game requires. There are so many variables that become multiplied and compounded depending on each circumstance, finances, contracts, player ratings, the upcoming free agent pool, the talent pipeline in the minors, management and personnel philosophies, and on and on.

Chess moves and strategies can be reduced to a simpler mathematical problem using a more finite set of variables. There are only so many squares and pieces on the board, and there are only so many possible moves by you or your opponent within the rules. In essence, managing a baseball franchise and making complex personnel moves involves far more squares and game pieces and far more options and mathematical possibilities. Then there's the issue of actually coding the AI to consider all of those, make reasonable decisions, and do so consistently and correctly.

This is why we're probably stuck with much dumber AI in sports simulations for the foreseeable future. Getting to something much more powerful and sophisticated is probably beyond the coding capabilities, resources and budget of any game developer. But you can probably make incremental improvements and do things that are more immediately feasible, such as at least getting the AI to "understand" that spending all of its available money for free agents, extensions and team options on just one or two players is not a smart strategy. It can potentially be coded in some way to try to enable more transactions with that money.

[CBeisbol]

Originally Posted by Charlie Hough

Theoretically it's possible to code AI in a way that would allow it to plan ahead and think in sophisticated ways like a human GM can, but it would be staggeringly complex and would probably require far more coding and computational power than a chess game requires. There are so many variables that become multiplied and compounded depending on each circumstance, finances, contracts, player ratings, the upcoming free agent pool, the talent pipeline in the minors, management and personnel philosophies, and on and on.

Of course it's complex because there are infinite possibilities. Just peak into my mind when trying decide whether to deal one random minor leaguer or not.

But it wouldn't need to do all that. It could have some basic assumptions built in and run a few sims to see if it's better to keep Ahmed or throw a bunch of resources at Lindor.

To see if things

such as...spending all of its available money for free agents, extensions and team options on just one or two players is not a smart strategy.

Is true or not.

[Timofmars]

Scouting accuracy and also the AI evaluation settings can have an effect too. They may have a low opinion of the player they are trading away because their scouting is inaccurate. The GM has different valuations of players that can be more based on current year statistics or previous years instead of scouted ratings, or the GM can have trade preferences for veterans over rookies, or for speed over power, or contact over OBP. And they can also consider if they have enough backups or prospects for a position and are more willing to give them up for some other position they think they need.

[BillyBall]

Just a follow up to Lindor's fate.

He ended up going FA and signed a 8 year deal with the Giants at about $27M per year. His demand price at FA was $43M per year for 10 years, but he was one of the first off the board and came down a lot on his agreement.

So, in the end the D-Backs as an out of contention team paid $6M for a couple months of Lindor action and lost a very good and cheap young player also.

[Syd Thrift]

Originally Posted by CBeisbol

I'm not a programmer
But isn't the advantage of computers their ability to perform lots of calculations.

That's why computers are good at chess. They analyze lots of moves, and lots of moves ahead.

Perhaps it's my lack of knowledge in the subject area, but it seems to me, the AI should (or could) excel at this.

It took people decades to create an AI that could play chess well, and it’s been even slower going for other games because chess has with it the intrinsic advantage that you have a series of symmetrical moves that you can plan ahead with. Other games like Go have proven much, much harder (although I believe that there has been some recent breakthrough in that).

On top of that, I strongly suspect that the *last* thing people want is a neural network style AI that records the result of every choice it makes and responds. For one thing they tend to take several hundred or thousand iterations to map things out, which means that having one that isn’t just guessing would require you to have, like, everyone connect to a server and just live with the crappy AI for a couple of months.

Second, I doubt people actually want that kind of optimized AI. They want, instead, an AI that acts the way humans act and makes the kinds of mistakes humans make. What if, like, the AI determined that the ideal pitching staff was 15 men with nothing but openers? Like, something so extreme that it’d decide that the added advantage of never using a tired pitcher was worth occasionally having to put a pitcher in the field in case of an injury? What if the AI noticed a game generated market inefficiency involving catchers and gap power that caused it to scoop up and, in human terms, grossly overvalue any catcher who hits doubles (I raise that because I distinctly remember OOTP 4 I think having that issue)?

Maybe some people would find this interesting but I suspect that the vast majority of players would cry foul. And so, Markus and company have little choice but to hard code the AI, which means the limits are the imagination and/or ability to effectively expand on code of one or two developers, not the black box of the sorts of AI you see in movies.

[CBeisbol]

Originally Posted by Syd Thrift

Second, I doubt people actually want that kind of optimized AI. They want, instead, an AI that acts the way humans act and makes the kinds of mistakes humans make. What if, like, the AI determined that the ideal pitching staff was 15 men with nothing but openers? Like, something so extreme that it’d decide that the added advantage of never using a tired pitcher was worth occasionally having to put a pitcher in the field in case of an injury? What if the AI noticed a game generated market inefficiency involving catchers and gap power that caused it to scoop up and, in human terms, grossly overvalue any catcher who hits doubles (I raise that because I distinctly remember OOTP 4 I think having that issue)?

Maybe some people would find this interesting

I would be one of those people.

And so, Markus and company have little choice but to hard code the AI, which means the limits are the imagination and/or ability to effectively expand on code of one or two developers, not the black box of the sorts of AI you see in movies.

What about just doing some simple tests? Hard-coding some basic guidelines (to prevent the above) and then looking at the short and long term outcomes of doing this move, or not doing this move?

Deciding which moves to check (and which not to check), and having the ability to do the checks without tying up the game doing full sims for each one, would be, I would think, the issues.

[Syd Thrift]

Originally Posted by CBeisbol

I would be one of those people.

Yes, I figured, but like I said the overwhelming majority of people would be against it, and if it were an opt-in type situation, very few would opt in, especially after it started producing the typically horrific results that neural network AIs tend to produce in the early stages. Especially when the final "payoff" is... weird, inexplicable crap (see below).

What about just doing some simple tests? Hard-coding some basic guidelines (to prevent the above) and then looking at the short and long term outcomes of doing this move, or not doing this move?

Neural networks are not AFAIK really good at looking at short and long term outcomes like this. I'm grossly oversimplifying this, I'm sure - I don't program AI - but the basics of a neural network is that you give the AI a set of data, you allow it to fire up, say, 1000 different versions of itself to make a series of discrete decisions - ideally the same decisions, which would make it that much harder if you set it up for people to play against, and if you tried to just sim the game against the current AI all it would be doing is figuring out how to exploit said AI - even using that as a basis for a 2nd or 3rd generation to play against players would probably come with some wonky results that you'd have to take a generation or more to untrain the AI.

It's not very easy to get one's head around. In fact, kind of the point of neural networks / machine learning is that you're not supposed to get your head around how exactly it comes to a given conclusion at all - a neural network is basically a "black box" where you put data in, you watch the outcomes of the data, you tell it what the best outcomes it came up with are so that it can use those for its next generation of power, and so on and so forth.

I'm not seeing how that would be anything close to a stand-in for the current AI. At best, it'd suck for a while and then finally figure out that the ideal way of doing things is the way GMs have been doing them all along, which strikes me as very unlikely. At worst, it'd come up with very alien methodologies. And through it all, all the dev team would be able to say to the "oh my god what did you do to the AI it was bad before but it's terrible now and still sucks!!!" crowd is, "be patient and also maybe whatever it's doing to your game is the real optimal way baseball GMing should be done".

The other, bigger issue is that OOTP would need to bring in a dev who knows how to work with neural networks / machine learning. That's a pretty specialized skill and people who are good at it make a crap-ton of money. It's at the point now to where even if one of the current devs were to train themselves up in machine learning to the point of confidence, they could just go to like Google or somewhere similar and make scads more than they ever could at OOTPDev. Maybe in 10 or 20 years when the market has died down they can reconsider.

[CBeisbol]

Originally Posted by Syd Thrift

Yes, I figured, but like I said the overwhelming majority of people would be against it, and if it were an opt-in type situation, very few would opt in, especially after it started producing the typically horrific results that neural network AIs tend to produce in the early stages. Especially when the final "payoff" is... weird, inexplicable crap (see below).

One entity's "weird, inexplicable crap" is another's "wow! Awesome"

Neural networks

You're the only one talking about a neural network

I'm just saying, the current AI could have a binary yes/no option, say, for a trade offered to it. Should the DBacks trade some prospect for free agent to be Lindor. Yes or no.

It could then sim some number of seasons and compare the short and long term results to compare making the move and not making the move. Then use that info to decide whether or not to make the move. That's it (it almost certainly being pretty difficult).