An open letter to Markus regarding AI evaluations/decision making

[red95vette]

I'll might get flamed by some people for this post -- so go ahead if you feel the need and take your shots. But, I'm hoping that Markus at least gets to read this and maybe it will strike a chord, turn on a lightbulb, or whatever and ultimately make OOTP better as a result.

This is my second year of evaluating OOTP. I trialed it last spring and did not purchase then. I'm nearing the end of a trial on OOTP8 and have not made a firm decision yet, but am leaning towards again not purchasing.

But, it pains me. Because the game looks to be REALLY fantastic in so many areas. However, the thing that still holds it back for me personally is the computer AI specifically with regards to trade AI and roster management (or even more accurately "organization" management).

Multiplayer-human leagues versus solo play:
I don't have the time required to participate in online leagues. SO, I really look at the solo game aspect of OOTP when evaluating whether it's right for me. And although I've tried playing through the demo with self-imposed "house rules", it really makes it much less fun, and certainly not "immersive" in the sense that I am really running a franchise when I have to always check to see if the transactions the computer GMs are making are reasonable or not.

So please, anyone replying with a knee-jerk response to "just play multi-player" can just save their breath (typing) as that's not what I am looking for.

I would think Markus intends for OOTP to be a rewarding experience for solo players as well. Otherwise, why even bother with features geared towards that mode? If he doesn't then that's fine and he should say as much as I'll stop bothering to check out each demo as it's released.

What are humans good at, what are computers good at?
I mean let's face it, we humans have the ability to think about things and arrive at decisions intuitively, using all sorts of intangible information. We also have the power to use adaptive reasoning and when something isn't working to our satisfaction, try another approach.

But, the computers main strength is the power of computation. If I can outsmart the computer GMs with intuitive thinking, that's OK. That's the advantage we human players have and it's what makes games like this fun -- trying to use our intuitive thinking to gain the upper hand.

But anything related to number crunching... the computer should be able to kick my royal butt all over the face of the planet right? That's where we humans give up the advantage to the computer.

It's all about the coded rules:
Well I happen to program for a living -- control systems which automate the operation of manufacturing processes. These systems use "AI" to make "decisions". Yes, human intervention is sometimes required (most notably when process devices fail, as I have yet to find machines that will repair their own physical components). But the point is 99% of the time, these systems operate autonomously at a level of decision making efficiency equivilent or greater than that of a human operator. Mainly because they are able to process so much information so quickly. As long as their "decision-making" rules are complete and accurate, they do an excellent job. If the rules are flawed, the results can (and usually will be) flawed.

An example with respect to OOTP trade AI:
I think anyone who keeps an open mind and reads the following example carefully will follow along and agree. Hopefully Markus will as well.

I had a thought about something and tried an experiment that confimed what I suspected about how OOTP AI "evaluates" trades (and probably in-game resources in general).

I opened up a trade window in the game I'm playing and picked a team at random. I offered to trade them $2 for $1 in return. The response from the AI was "that's not a fair deal, you have to offer us some more." I increased the amount I offered -- all the way up to $500,000 in fact. I got the same message each time. Trade rejected....

My trading difficulty is set to "very hard", but that's a moot point really. (By the way, reducing it to Hard or Average didn't get the trade done either. Setting it to "Easy", the computer would FINALLY let me give away a cool half mil -- but just barely).

There is absolutely no circumstance under which this trade would not be advantageous to the computer. And even at just $2 offered to $1 requested the computer should have jumped on the offer. Here's why.

Risk versus Reward and ROI (Return on Investment)
I suspect I know EXACTLY why the computer did not want to accept the trade. I think it has been coded to only trade when it thinks the advantage it is getting exceeds a certain amount. The specific amount is no doubt tied to trade difficultly level in AI options.

This approach is flawed though, because it does not take into account risk/reward and ROI. I think this is why many OOTP players have posted one of the following three opinions on these forums:
(1) The trading AI logic is TOO HARD
(2) The trading AI logic is TOO EASY
(3) The trading AI logic is JUST RIGHT

Funny thing is all three can be true.

--In this cash for cash scenario, the computer has no risk whatsoever, so it should jump on the trade even if it's only $2 offered for $1 given.

--In a scrub for scrub trade (or scrub for cash), the risk is small as there are usually plenty of replacement players available for minor league (0 cost) contracts sitting in free agency. Again I think ROI principles should be applied. If it "thinks" it's player is worth $50K, then it should be willing to accept "assets" (players or cash) that exceed that value. Given the small risk, it should want a profit, but that should be commensurate with the value it is giving as well as AI trading difficultly level. Maybe something like:

Very Easy = 0% ROI
Easy = 5% ROI
Average = 12% ROI
Hard = 20% ROI
Very Hard = 30% ROI

So the computer would want assets worth at least $50K on easy level (no advantage), and $65K on very hard.

--In a trade involving players better than replacement level and/or including prospects, the risk is quite high. Very good player's free-agent "value" can be calculated, but there is no guarantee that such FAs actually exist to replace these guys at the present moment. There is also the risk that any available FAs will ask for more money than they're worth or will sign with another team. It's because of these factors that I would not mind at all if the computer GMs marked certain key personnel as "untouchable" at any price if they are starters and no viable replacements with similar skill levels exist already within the computer GMs organization. (I don't think I've ever seen a player in OOTP that couldn't be "bought" if the price was high enough, even if that player was a key starter and all that was being offered was prospects, albeit good ones). It also doesn't seem to matter to the computer if the "good prospects" being offered are at positions already talent rich in the organization.

Furthermore, even if the computer GM is in "rebuilding" mode and is trading away established veterans for prospects that it calculates to be of somewhat "equal" value, there is still high risk that the prospect will not turn out. Or stated another way, the "values" of the assets are so high and the risks so high that the computer should look for way more of an advantage in "real-dollar" terms than it presently does.

SO maybe for a high risk trade the ROIs are something like this:

V. Easy 5%
Easy 10%
Average 20%
Hard 35%
V Hard 50%

But even further since the "value" of the assets on each side is so much, the advantage required by the computer (since it should calculate it on a %ROI basis) should be much higher as well. Say the user asks the computer GM for his ace starter (valued at $15M as an asset). The computer should want about $18M in return value on average and $22.5M on V. Hard.

Maybe this value is adjusted if the computer is trading away players with "overvalued" contracts... but I would suggest that only be the case if they are in bonifide rebuilding mode or if suitable FA replacements are plentiful (which by definition in a way would make this player much LESS valuable to begin with since the "average" player in that league would be much better).

Conclusions

So in summary, I believe the following statements are correct.

-- The computer seems to not recognize zero or minimal risk / high possible gain situations when trading and take advantage of those when they are presented to it. This is much to it's own disadvantage.

-- The computer seems to evaluate "advantage" on non risk-adjusted basis therefore, on very high risks trades it does not seek ENOUGH of an advantage. Again, this is much to it's own disadvantage.

-- The computer seems to allow all of it's organizational assets to be "for sale" at the right price, rather than consider front-line starters untouchable (especially when they are much better than replacements available with the organization already).

-- The computer should look at the assets being offered and further devalue them (even to the point of zero or negative value) if they are for positions in which there is ample talent throughout the computer's organization already. (Negative value obviously only being applicable if assets offered have a ML contract).

I think these ideas alone would go a long ways towards making the computer GM AI more competative and possibly even spill over to other AI evaluations such as FA signings and draft picking.

I'm hoping to be a satisfied OOTP player/customer at some point in the future. The program is already great in so many ways, If it were just a little bit "smarter" when it came to simulating GMs I'd be completely immersed and hooked with no hope of recovery. I offer these comments as constructive criticism. I believe alot of potential solo players probably feel similarly but may not take the time to explain why they demo but don't ultimately purchase. So here's hoping this won't fall on deaf ears.

[johnnybourbon]

you obviously put a lot of time, energy and research into this post. thank you for that. I have noticed flaws in the AI trading, but havent gotten into specifics, for I figure I'll just crank up the trading difficulty and live with it when the AI doesnt let me fleece its prospects out from under it anymore.
I believe theres a recent thead discussing research done into "make this work" vs. "shop a player", Can someone link to it in this thread?

[RonCo]

This is among the reasons I will only play for real against other humans. You're just not going to get a complex AI interaction like this game needs without a whole lotta design and coding resources applied to it. My opinion, of course. YMMV.

[RchW]

Nice post. No flames here.

Based on the depth of your analysis, it would be interesting if you could somehow evaluate human/human trades in an online league.

[endgame]

Very interesting thoughts, rv. Thanks for taking the time to articulate both your method and your findings. This has inspired me to look into this arena even further. That, in and of itself, is a good result from your post. I think we all hope that in the versions to come over the years, the AI will be brought more and more to 'life'.

[red95vette]

Thanks for the positive reaction, guys.

Originally Posted by RonCo

This is among the reasons I will only play for real against other humans. You're just not going to get a complex AI interaction like this game needs without a whole lotta design and coding resources applied to it. My opinion, of course. YMMV.

I would love to play in an all-human league. I think it would be a blast. The problem I have is my gaming time is very irregular. I have periods (between work projects) where I can game for 2-3 hours a day easy. Those times would be ideal for online league play. Unfortunately, I also go through periods where I don't have time to do ANYTHING but work, eat, and sleep (and the latter two are sometimes optional). So, I can't in good faith sign up for a league when I know about half the time I can't give it any attention.

Anyway, I don't think the AI interaction is as "complex" as you might think. Afterall, OOTP already models most of the things I described. It "values" players obviously to be able to determine how much money players ask for and also to determine what their worth is for the current trading system. It also tries to gain an advantage on any given trade as illustrated by the fact that it wants more than $500K advantage before it will do a deal on a cash-for-cash basis. Really, my OP only proposed a couple of additional things with regards to trading specifically.

(1) Adjust the advantage sought by the AI based on the "riskiness" of the transaction. Cash is cash, it won't change. Players can and will change. The "riskiness" of a player asset is correlated with how high his current/potential ratings are -- the higher above replacement value, the more "risk" since changes at that end of the ratings spectrum equate to drastic player valuation changes.

(2) No matter what the "value" situation looks like for any transaction, consider the entire organization's player situation. Never trade away the front-line starters just because the "value advantage" offered is large. Mark those players as "untouchable", unless a suitable replacement player is already in the organization (or possibly one is offered in trade). The exception being obviously if the team is no longer in the playoff race and it is intentionally trying to dump salary or unload a bad morale player.

You know it's a funny thing because I often seem to run across the opinion that "AI" is complex and impossible to do well, et cetera whether that is with regards to a sports sim game or automated manufacturing plants. The truth is that AI programming is NO HARDER to do right than any other sort of programming. OOTP is already programmed with some pretty good player/free-agent behavior AI that seems damn good to me. And I know the OOTP development group is a talented bunch because the rest of the OOTP programmed universe is AMAZING. Really all that is necessary for good AI programming is to sit down and think of all the ways you want the computer to "act", develop a plan/model that encompasses that behavior and then code it. The problem I've seen in my industry is when AI is coded in a hodge-podge fashion instead of to an overall plan. The problem with that approach is that often individual bits of logic can start to conflict with others that were put in -- and then adding "band-aid" logic to fix niggly problems usually just end up making them worse or breaking something else. It's like any other good programming practice: just develop a plan and then program to it. If in testing something comes up, it's either a bug or a flaw in the plan. Obviously bugs get fixed and flaws in a plan get rethought. On some very rare occasions you might find it "impossible" for AI to evaluate some particular situation correctly. In such instances, recognize that in your plan, and develop an approach to ensure the situation won't come up. But these instances are quite rare when the whole situation is modeled out in a masterplan.

Let's be fair. I didn't say I wanted the AI to act in such a way that it is impossible to tell that you're dealing with a computer GM in the game rather than a human. That expectation is not fair. However, there is plenty of numerical data available and the computer can do a GREAT job of calculating all the "values" of assets involved in any transaction and it also can do a great job of calculating the potential risk associated with any transaction (when coded to a master plan that defines those risks). Using this data and having "perfect" knowledge of all the in-game data available, the computer should always be able to make smart informed "AI decisions" about any facet of gameplay. Whether those AI decisions pan out in the long run then is the result of the SIM and sometimes other random factors. In fact, when properly coded the only way the AI would not be able to come out with the best end of any transaction is if it was purposely programmed to use uncertain information, such as the human uses when scouting is on.

It's a bit of another tangent, but I've noticed someone else posting about (paraphrasing as I recall it) the fact that the AI would do an inaugural draft and then immedaitely after it's complete (without simming a day) trade it's early round draft choices for a human's late round draft choices. This quite simply should not be posssible. It distinctly points to the fact that there MUST Be multiple ways the AI is evaluating players, agreed? Why is this happening? It's even worse when you consider the AI thinks it's getting a bargin on that trade since it's trying to gain an advantage. So that infers that it not only passed up a more valuable asset during the draft, but it now also believes it is significantly more valuable.

Anyway, the simple point is remove the obvious deficencies from the AI like we've discussed in this thread and the computer immediately becomes a much more challenging opponent -- even if it lacks advanced ad-hoc reasoning skills of humans. Afterall, it can kick our butts on the "valuation" side of the equation if it is programmed to do so.

[RonCo]

Originally Posted by red95vette

I would love to play in an all-human league. I think it would be a blast. The problem I have is my gaming time is very irregular. I have periods (between work projects) where I can game for 2-3 hours a day easy. Those times would be ideal for online league play. Unfortunately, I also go through periods where I don't have time to do ANYTHING but work, eat, and sleep (and the latter two are sometimes optional). So, I can't in good faith sign up for a league when I know about half the time I can't give it any attention.

You can probably find a league where that level of involvement would work.

Anyway, I don't think the AI interaction is as "complex" as you might think. Afterall, OOTP already models most of the things I described. It "values" players obviously to be able to determine how much money players ask for and also to determine what their worth is for the current trading system. It also tries to gain an advantage on any given trade as illustrated by the fact that it wants more than $500K advantage before it will do a deal on a cash-for-cash basis.

I dunno. I've studied the game for a pretty long time, and I'll readily admit I can never figure out what the AI is doing. But I would suggest that the fact that the AI won't trade $1 for $2 as being a pretty big bug.

When I say "complexity" I mean from the real-world point of view. The things that make a trade happen in real life, for example, are often a very complex balance of skill and business elements that I would be honestly pretty surprised if a single game developer was really capable of doing in a consistently excellent fashion. OOTP certainly models a lot of the right thing, but having a model of a thing, and having that model actually work right are two different things.

I mean, if we can't get something like "distance to the LF fence" to affect the likelihood of a HR occurring, I'm having a hard time seeing us getting the AI to do good things all the time.

Really, my OP only proposed a couple of additional things with regards to trading specifically.

(1) Adjust the advantage sought by the AI based on the "riskiness" of the transaction. Cash is cash, it won't change. Players can and will change. The "riskiness" of a player asset is correlated with how high his current/potential ratings are -- the higher above replacement value, the more "risk" since changes at that end of the ratings spectrum equate to drastic player valuation changes.

(2) No matter what the "value" situation looks like for any transaction, consider the entire organization's player situation. Never trade away the front-line starters just because the "value advantage" offered is large. Mark those players as "untouchable", unless a suitable replacement player is already in the organization (or possibly one is offered in trade). The exception being obviously if the team is no longer in the playoff race and it is intentionally trying to dump salary or unload a bad morale player.

You need to write a lot more use-cases that this.

You know it's a funny thing because I often seem to run across the opinion that "AI" is complex and impossible to do well, et cetera whether that is with regards to a sports sim game or automated manufacturing plants. The truth is that AI programming is NO HARDER to do right than any other sort of programming.

I've written code since I was 12. And I'm a lot older than that now. Programming is pretty easy. Design is way more complex.

OOTP is already programmed with some pretty good player/free-agent behavior AI that seems damn good to me. And I know the OOTP development group is a talented bunch because the rest of the OOTP programmed universe is AMAZING. Really all that is necessary for good AI programming is to sit down and think of all the ways you want the computer to "act", develop a plan/model that encompasses that behavior and then code it. The problem I've seen in my industry is when AI is coded in a hodge-podge fashion instead of to an overall plan.

That _is_ the root of most software problems. It's also what can make AI so difficult to "fix." I suggest the existence of so many legacy AI issues is indicative of either (a) this type of development model making the process "complex", or (b) the lack of desire on Markus's part to fix them. Since I know for a fact Markus has a high degree of interest in fixing problems, I can only fall back on the idea that the solution is too complex for him to have gotten some of these models right...or I guess that he's just not sat down to design them...though since his customers have been complaining about things like "sign and release" for years I would expect he's gone to the drawing board a number of times to try to design such a solution.

The problem with that approach is that often individual bits of logic can start to conflict with others that were put in -- and then adding "band-aid" logic to fix niggly problems usually just end up making them worse or breaking something else.

Hmmm....

It's like any other good programming practice: just develop a plan and then program to it. If in testing something comes up, it's either a bug or a flaw in the plan. Obviously bugs get fixed and flaws in a plan get rethought. On some very rare occasions you might find it "impossible" for AI to evaluate some particular situation correctly. In such instances, recognize that in your plan, and develop an approach to ensure the situation won't come up. But these instances are quite rare when the whole situation is modeled out in a masterplan.

Masterplan...there's an interesting word.

Let's be fair. I didn't say I wanted the AI to act in such a way that it is impossible to tell that you're dealing with a computer GM in the game rather than a human. That expectation is not fair. However, there is plenty of numerical data available and the computer can do a GREAT job of calculating all the "values" of assets involved in any transaction and it also can do a great job of calculating the potential risk associated with any transaction (when coded to a master plan that defines those risks). Using this data and having "perfect" knowledge of all the in-game data available, the computer should always be able to make smart informed "AI decisions" about any facet of gameplay. Whether those AI decisions pan out in the long run then is the result of the SIM and sometimes other random factors. In fact, when properly coded the only way the AI would not be able to come out with the best end of any transaction is if it was purposely programmed to use uncertain information, such as the human uses when scouting is on.

Risk analysis is certainly an area that I would guess could be improved. In addition, some element of game theory and/or understanding your opponent's need or hunger to make any deal could be included to make things even smoother.

It's a bit of another tangent, but I've noticed someone else posting about (paraphrasing as I recall it) the fact that the AI would do an inaugural draft and then immedaitely after it's complete (without simming a day) trade it's early round draft choices for a human's late round draft choices. This quite simply should not be posssible. It distinctly points to the fact that there MUST Be multiple ways the AI is evaluating players, agreed?

I would phrase it something more in the line of "there are inconsistent methods of analyzing players at different times...or, there is no team-by-team 'masterplan'" to steal a word.

Why is this happening? It's even worse when you consider the AI thinks it's getting a bargin on that trade since it's trying to gain an advantage. So that infers that it not only passed up a more valuable asset during the draft, but it now also believes it is significantly more valuable.

The only thing we can say with any sense of certainty is that something is wrong.

Anyway, the simple point is remove the obvious deficencies from the AI like we've discussed in this thread and the computer immediately becomes a much more challenging opponent -- even if it lacks advanced ad-hoc reasoning skills of humans. Afterall, it can kick our butts on the "valuation" side of the equation if it is programmed to do so.

The game is going into version 9. These "obvious deficiencies" have been being discussed since I came into the environment (in version 5), and were considered legacy issues then. If AI development is not as complex as I think it is, you're going to have to explain to me how Markus has been unable to fix them, despite every release including "AI improvements" in the feature list.

That said, I've always been in Markus's camp here. I don't hold him accountable for much more than rudimentary AI behavior.

[PSUColonel]

Well, I too must admit that many of the roster management issues have been in almost all incarnations of the game. That being said, I am of the opinion that Markus would like to fix these issues, and that it very may well be that he simply can't. (for whatever reason)

[red95vette]

Originally Posted by RonCo

The game is going into version 9. These "obvious deficiencies" have been being discussed since I came into the environment (in version 5), and were considered legacy issues then. If AI development is not as complex as I think it is, you're going to have to explain to me how Markus has been unable to fix them, despite every release including "AI improvements" in the feature list.

That said, I've always been in Markus's camp here. I don't hold him accountable for much more than rudimentary AI behavior.

I appreciate all of your responses they are well stated and thought out.

And, on whole I can say we see eye-to-eye on this more than not. Your last question is somewhat loaded and I won't venture to answer as obviously I have no insight into what has been the development process/evolution of the product.

However, based on impressions I've formed by reading many, many threads in these forums, I will postulate a couple of indirect theories that lead me to believe such AI issues haven't been very high priority in the past, even if they've been looked at numerous times, simply because:

1. Lots of players profess to be online league players only. AI is much less important when all GMs are humans.
2. I've seen some solo players on here state the old "just use house rules" and "don't cheat/exploit the AI" mantras.

At the end of the day, for a prospective solo only player though, with a very competitive streak, #2 rings hollow. It's a fine line between "cheating the AI" and "making a shrewd GM move".

For me personally, all immersiveness is destroyed if I have to worry about whether I am being "fair" to the computer GMs in a trade or by drafting the best talent available when their teams may really need that talent more to be competitive. Like I stated before, it should have me completely whipped when it comes to access to data and computational power. I should be looking to gain an advantage elsewhere.

Maybe the game is just too complex at this stage to fix existing AI issues without a total rewrite of that code, something the development team is likely not keen to do. If so, that answers my question about whether I need to keep demo-ing the game when subsequent versions are released.

[RonCo]

I'm certain that if we were sitting together and sharing our favorite soft drinks we would be of a like mind before long.

With regard to your (1), I have played a lot of solo OOTP in every version as a tester. I don't play solo for fun because, well, for me it's no fun to beat the AI. Other's opinion may vary. But quite honestly, if I work at it at all, I'll almost certain to be in the World Series within 3-5 seasons whenever I take on a team in a solo league...unless I just completely stack the deck against me...and even then it's 50-50. Perhaps I'm overstating it a little, but not much.

[RonCo]

Even going to human only sim-leagues, the in-game AI still gets you. My run in the playoffs last season was brought to an AI-enhanced early end when I clinched my division title way early, so "rested" my top two starters the last sim, bringing two minor leaguers up to start. So, down several runs in the last game of the season and heading into the playoffs, the AI takes my Cy-Young superstar starter who I'm "resting" and throws him 6 innings in relief in a losing cause, thereby limiting him to one start in my 5-game series, which I lost 3-2 after being up 2 games to 1.

This seems like an incredibly stupid thing for an AI to do. I mean, not even Dusty Baker would go that far. Why does this happen in v7/8? I dunno. I don't think Markus purposefully coded an algorithm that wore my best pitcher out in a low value situation. I suggest it's because the process of writing a comprehensive AI that can handle all baseball situations is a huge amount of work, and requires incorporation of a high degree of complex thinking. It's a bug of omission.

[red95vette]

Originally Posted by RonCo

I'm certain that if we were sitting together and sharing our favorite soft drinks we would be of a like mind before long.

With regard to your (1), I have played a lot of solo OOTP in every version as a tester. I don't play solo for fun because, well, for me it's no fun to beat the AI. Other's opinion may vary. But quite honestly, if I work at it at all, I'll almost certain to be in the World Series within 3-5 seasons whenever I take on a team in a solo league...unless I just completely stack the deck against me...and even then it's 50-50. Perhaps I'm overstating it a little, but not much.

I can see this as the case for sure. In my limited time with it, I found it to be pretty easy, even following suggestions to turn off actuals and use potentials only (and no overall). Maybe I was lucky... but being a N00b, I expected a bit more challenge at least from the get go.

Originally Posted by RonCo

Even going to human only sim-leagues, the in-game AI still gets you.

Ouch, that's a painful story. The "bug of omission" is dead on though. At least in solo play, you can control your team in that situation. In a SIM you are completely at the mercy of Markus. Must have been something you said.

[Le Grande Orange]

Originally Posted by RonCo

My run in the playoffs last season was brought to an AI-enhanced early end when I clinched my division title way early, so "rested" my top two starters the last sim, bringing two minor leaguers up to start. So, down several runs in the last game of the season and heading into the playoffs, the AI takes my Cy-Young superstar starter who I'm "resting" and throws him 6 innings in relief in a losing cause, thereby limiting him to one start in my 5-game series, which I lost 3-2 after being up 2 games to 1.

This seems like an incredibly stupid thing for an AI to do. I mean, not even Dusty Baker would go that far. Why does this happen in v7/8? I dunno.

If I had to guess, it's because the AI doesn't "know" there is a more important goal to strive for, that is, the post-season. It just thinks "in the now" so to speak; you're losing that last game, and so it does what it thinks necessary to try and win that game. It can't see the longer term implication of that, namely, using that pitcher in that last game jeopardizes the upcoming post-season playoff series.

[Syd Thrift]

Originally Posted by RonCo

This is among the reasons I will only play for real against other humans. You're just not going to get a complex AI interaction like this game needs without a whole lotta design and coding resources applied to it. My opinion, of course. YMMV.

", only it's why I play in God mode.

[Syd Thrift]

I think the larger issue here is, how do you keep this all under the hood? As much as I emphasize with RonCo's example, how else is the computer supposed to handle the situation? It sees it has an empty bullpen (or not; maybe the particular role the AI would use doesn't have any rested players) and a really, really good guy not listed with a role. 6 innings is a bit much, but if this same guy was a 6th starter who was only on the roster because rosters expanded would we bat an eye at this?

I think what RonCo would have probably preferred was a "shut this guy down" button. But that button is just going to make a game that is already getting people angry at it for being too complicated even moreso. So does Markus input some code resembling Italian pasta to the tune of "if pitcher X is listed as a starter but doesn't have a role and is 140% better than the league average, do not use him"? That might resolve RonCo's initial issue but I don't think that it resolves the overall thing: that the AI is artificial, not human.

If I were a smarter person with more time on my hands, I would track the transactions and strategies used by people in online leagues and develop an AI off of that. Of course, it would still make dunderheaded moves, but a perfect AI that never, ever makes bad decisions is not what I (at least) want either.

A secondary thing... one thing I'd like the AI to do is game the "Win Now/Rebuild" thing a lot, lot harder. It's more challenging to play in a league against a couple of powerhouses and a few patsies than it is to play against a bunch of teams that oscillate between pretty bad and pretty good. It's also more realistic to baseball; up until free agency, you regularly, decade after decade, had 2 or 3 teams up at the top of each league beating up on everybody else. This gets more and more pronounced the further back you go. Although I do see OOTP teams occasionally get into the doldrums and not climb out of the second division for, say, 30 years, I see this happen to one team in 16, not 6.

Also... while I get the point about AI... I don't think it's entirely realistic to have teams engage in wholly rational RvR reviews every time it makes a trade. I guess that underneath the hood this should (and probably, to some extent, is) be a way to evaluate potential trades/free agent signings/etc. but there are other bits to consider:

- Some GMs ought to have wildly different criteria for judging talent than others. This happens to some degree right now, but I want one guy who looks at guys based entirely on their tools, dammit. Each GM ought to have an ability or two that he prizes above all else, as well as a couple attributes he really doesn't care about. Billy Beane might rank Patience highly and rank all the "little skills" like Baserunning and Bunting low (and would probably be ineffective if placed in the dead ball era). George Steinbrenner, in the days when he was so intrusive he was basically the Yanks' GM, might rank Popularity very high but not care much at all for some of the character ratings. In the end, the guy who is highly evaluating skills that nobody else is looking at is probably the person who has an edge.

- Penalties for "moneyball"? I think there should be the potential for a player who is used to playing in the 3-hole his entire career to get really, really hot steaming mad if he bats leadoff, no matter what his OBP is. I know that some here don't want morale to influence on-field play that much, and I'm fine with not changing that much from now, but... what about if one players' anger is contagious? Right now, I don't think it is. The causes of low morale can affect an entire team at once, so a bad team especially gets a lot of sad players, but IRL if you have one guy in the clubhouse who isn't happy, he's going to whine to everybody else, and that's going to have a negative effect (even if it's only because everyone else wants the guy to be the hell quiet). Let's put this into the game! For added baseball goodness, maybe in extreme circumstances there could be clubhouse crises over a given player, causing half the club to gain morale when he's traded off and have to lose it.

- Another thing that OOTP pulled from WWSM that I want to see implemented more thoroughly in the game: role on the team. I think this should be part of contract negotiations. I can buy a player only choosing to play for teams who tell him they will start him and so on, and even believing a team that's lying through its teeth when it says this, but... if you lie to a guy and sit him on the bench, not only should he be unhappy about this (see above comment), you as a GM should get "bad boy" points a la Europa Universalis that cause future FAs, guys at the end of their contracts, and maybe even GMs to force you to give them better than average deals because they know that you're more likely to pull the rug out from under them.

- One last axis that should be considered when evaluating... some GMs will spend millions in free agency but won't drop a dime on guys in the draft. Right now, there is no signability factor in the draft. Maybe there should be. If there is, it shouldn't automatically come down to the richest teams spending more than the poorer teams. The Yankees are a great example of a team that, in years past, has not reached all that often for costly, high-reward guys. Costly, high-popularity guys maybe (Drew Henson comes to mind).

- Also - and this may open up a bigger can o' worms than ever thought possible - I'd like the option to turn off the draft and have teams sign kids right out of HS and college. Potentially unwieldy? Yes, of course. It was heavily unwieldy IRL, and continued to be even after the leagues added the draconian bonus baby rules (which, generally speaking, didn't do a whole lot other than burn out a few prospects who should have spent a couple years in the minors but couldn't) and didn't get better until the draft. Still, since it's start of baseball history...

[RonCo]

Originally Posted by red95vette

Ouch, that's a painful story. The "bug of omission" is dead on though. At least in solo play, you can control your team in that situation. In a SIM you are completely at the mercy of Markus. Must have been something you said.

You pick your poison. The actual "ideal" game would be one where you can play other humans, and control every managerial aspect of every game. But even I don't have enough time for that.

[RonCo]

Originally Posted by Syd Thrift

I think the larger issue here is, how do you keep this all under the hood? ...

Your whole post is essentially making my point. You've touched on a lot of inputs that go into any particular decision in a baseball world. To have a really good AI, you need to model something in every one of these spaces, and that model needs to work right. I just don't see a single programmer, especially one who is (most likely) not steeped in deep study of human psychology, will ever be able to make this kind of thing be truly robust. I have attempted to model human decision processes before. It's not a job for the faint of heart.

That OOTP works as well as it does is a marvel, really. It is something Markus can be pretty proud of. That said, I'll only actively play the game against other humans.

[SDtwin]

This thread is very interesting. I'm not gonna make a lengthy post with facts in it like you guys, but I really do like reading threads like this. It's people like this who help to make games better by providing realistic suggestion and not just complaining about problems. I think eventually Markus will get the AI working better or find someone who can help him do it. It would be awesome if he had another person like him to help code the game and offer different solutions for problems.