Sim Code Question For Jim G.

[Bonegavel]

Jim,

This may help others as well as myself, but could you explain a bit of why it is tough to code the AI to shop the FA market, or utilize things that we human players take for granted (simple things like putting the best guy in the line-up, etc)? Is it a trade-off for speed? IOW, allowing the AI to shop well would slow the game down to the point of turtledom. Or is the process so complex that it is impossible to model at this point in technology?

thanks

[sabotai]

IMO, it's hard because of the last thing you said. Just look at when we get a discussion of the top 5 running backs of all time going. There is a LOT of subjectivity that goes into what people think is a good player. It's hard to program subjectivity and AI, for as much as it sounds like it is advanced, I think it's pretty much still in its infantcy. Or at the most its terrible-twos.

To really model a good AI that takes into account a lot of factors and information and to have it come to a unique conclusion takes a good deal of memory and a LOT of processor power. (Which is why in most games, the AI seems stupid...most of the time for graphical games, they have to limit AI for the sake of the processor since it has to share processor power with the grapics. In sport sims, the reason is because you only have 1 or 2 people programming, and really great AI takes time to program.)

[Solecismic]

It's really that. How do you decide what players the AI should pursue? The combination of budgeting free agency dollars and getting the most bang for the buck is surprisingly complex.

Where it's most troublesome is deciding how the AI should pursue average starters. That's where the human player has had a decided advantage in the past. I'm working on a new theory on how to do this, and that's being tested right now.

I hope it works much better than the FOF4 system, because it will add a lot to the game if it does. I'm not sure I can come up with a better approach if it doesn't. Sabotai's right, really great AI can take years.

[Mac Howard]

>It's hard to program subjectivity and AI,

It is, but until you do it the game is woefully short on realism.

[Bonegavel]

Is it possible to record player actions, and try and create AI based off of that? Galactic Civilizations claims to do this, and I must admit that their AI is very interesting. I understand that no 2 rosters or situations will ever be the same, but there can be general guidelines to follow.

The game could dump choices of players to a file and based upon their overall score at the end of the year, it would seem that a pattern would emerge. Off the top of my head, there could be a flag that when the QB situation on a Team is X (this could be, age of starter/backup, rating of starter/backup, -anything really- or a combination thereof) this is what a consistently good player does, or bad player (if the coach is bad at that).

Again, I know that the variations are infinite, but there have to be certain and consistent actions that a player like QS does that a player like ME doesn't. This sort of approach would probably have to be written in from the first line of code, but, regardless of the complexity, there have to be patterns.

[sabotai]

Is it possible? I would think so. But I wouldn't call such a thing practical. Like you said, there are am infinite (not really, but enough) variations. You have to get the AI to examine countless numbers of seasons to even get close enough to an AI that would handle a team acceptibly, let long challenging to a human player.

And the time it would take to develope such an analysis AI would be too long for a single developer. Look at how people are getting on his case about not that many new features. If he spent the kind of energy this would take to improve the AI, improved AI would be all we get, and we'd have to wait a lot longer for it.

As for GalCiv, I'm not sure how that works. Sometimes it sounds like the AI can learn from the player files, and then the way some people describe it, all it is, is the computer using that single player's strat against you. Either way, doing it for a sports management sim would be tougher, IMO, because of what we want the outcomes to be. In GalCiv, it focuses more on the general strategy of the player instead of some of the fine details. It has to since every universe is unique. For a sports sim, it's the fine details that seperate a good AI from a great one, mainly because the "universe" is always the same.

(I'm not even sure if I understand what I just said....)

I have my ideas about how to create an AI to handle roster management for a football team (or any sport), but I haven't really begun to design and test it. One day....

[Kevin]

The situation described in Bonegavel's last paragraph could be potentially (and probably only partially) addressed by assigning the AI teams in FOF different computer GMs. These GMs would have different tendancies such that not all AI teams would work the same. It would also allow each team to act in a more consistent manner rather than making some random decisions that often conflict with each other.

Some GMs would favour free agents over draft picks. Some would play the hand they're dealt while others constantly look for fill ins from the free agents during a season. 4 or 5 of these variables could then help differentiate the teams from each other. These GMs would also periodically change teams.

I expect it would be a huge task to fine tune the effects of such a system though.

[Marc Vaughan]

Originally posted by Bonegavel
[B]Is it possible to record player actions, and try and create AI based off of that? Galactic Civilizations claims to do this, and I must admit that their AI is very interesting. I understand that no 2 rosters or situations will ever be the same, but there can be general guidelines to follow.

We do this to a certain extent in CM, the game 'learns' tactics and 'teaches' the better computer managers how to counter human tactical formations (assuming they have flaws - if they don't then it doesn't make much difference).

Players in the game also 'learn' about their managers personalities and adjust their reactions accordingly (this is more prevalent in CM03-04 than CM4 and NFG will have a lot more of this kind of thing reflected in the game).