Slitherine

quote:

ORIGINAL: berto

Again, whether we're talking time, disk space, whatever--I hope that there will be no (or as few as possible) arbitrary, fixed constraints on the AI. Let us decide for ourselves how much time we'll give the AI, and how much computer resources we'll devote to it.

quote:

ORIGINAL: a white rabbit

..no, computers are y/n, it's an electrical thing, the switch is either on , y, or off, n, there's no uncertantity state, maybe on/maybe off..

quote:

As to it's complexity beyond that of chess or iGo, a study on our own human CP analysis shows that isn't the case,

quote:

He doesn't have to analyse the whole map, just bits, one after another, as we do.

quote:

ORIGINAL: a white rabbit

..you're looking at it wrong, on the right magnification, a formation, say a corps is one 'thing', with a limited number of options,

quote:

..what toaw does now, is decide btn by btn, refering only to the pre-set objective track for that btn's formation,

quote:

ORIGINAL: golden delicious

quote:

ORIGINAL: L`zard

How many of you can prove to me that Your really alive and not an AI..........

I can do a good job. I've met seven other TOAW players in person, most of whom have posted here at one time or another. Meet any one of them and they can attest to my existence.

quote:

ORIGINAL: L`zard

I'll admit that meeting 'seven others' in person....WOWSERS ...you lucky bstd!

quote:

ORIGINAL: golden delicious

quote:

ORIGINAL: L`zard

I'll admit that meeting 'seven others' in person....WOWSERS ...you lucky bstd!

And just for fun, a short biography by TOAW. Forum names [thus];

2000ish: Peter Szabo at Waterloo Station
2001: Colin Wright, Richmond CA
2001: [Fidel Helms], NC (with Colin)
2001: Tim McBride, CO (with Colin)
2005: [Easily Confused] at the Cutty Sark, Greenwich
2006: [Specterx], [Viridomaros] simultaneously at Waterloo Station

quote:

ORIGINAL: Nemo69

That's quite a lot indeed.

quote:

I've only met with Viri in Liège in 2005.

quote:

ORIGINAL: golden delicious

quote:

ORIGINAL: a white rabbit

..you're looking at it wrong, on the right magnification, a formation, say a corps is one 'thing', with a limited number of options,

The options are finite, sure. But they number in the millions. Raise this to the power of the number of formations, and then add in the possible actions of the other side (bearing in mind sharply limited intelligence, whereas in Chess and Go intelligence is perfect).

quote:

..what toaw does now, is decide btn by btn, refering only to the pre-set objective track for that btn's formation,

This is more or less the problem. But the solution to this isn't a brute-force approach like you use for a Chess computer. You'd want a learning AI which tries to be more like a human player and less like an adding machine.

The idea I had would be for the player to design objective tracks independent of the formations. The AI would then judge how it was performing on each of these objective tracks relative to one another, and assign its formation to these tracks depending on the situation. So if the situation on one track is changing rapidly, that gets extra formations assigned to it.

quote:

ORIGINAL: Karri

Furthemore, the Supreme Command needs to 'know' how to give objectives(instead of designer implementing those objective tracks),

quote:

The key is information and how to use that information, and it should flow both ways. For example the 1st corps gives the 1st division the task of capturing the city of Kharkov. 1st division 'analyses' the situation and sends back a message that this is impossible without 2nd and 3rd division securing flanks. 1st corps then analyses the situation and decides if it can allow 2nd and 3rd division to do that.

quote:

ORIGINAL: golden delicious

quote:

ORIGINAL: a white rabbit

..you're looking at it wrong, on the right magnification, a formation, say a corps is one 'thing', with a limited number of options,

The options are finite, sure. But they number in the millions. Raise this to the power of the number of formations, and then add in the possible actions of the other side (bearing in mind sharply limited intelligence, whereas in Chess and Go intelligence is perfect).

quote:

..what toaw does now, is decide btn by btn, refering only to the pre-set objective track for that btn's formation,

This is more or less the problem. But the solution to this isn't a brute-force approach like you use for a Chess computer. You'd want a learning AI which tries to be more like a human player and less like an adding machine.

The idea I had would be for the player to design objective tracks independent of the formations. The AI would then judge how it was performing on each of these objective tracks relative to one another, and assign its formation to these tracks depending on the situation. So if the situation on one track is changing rapidly, that gets extra formations assigned to it.

quote:

ORIGINAL: Karri

quote:

ORIGINAL: golden delicious

quote:

ORIGINAL: a white rabbit

..you're looking at it wrong, on the right magnification, a formation, say a corps is one 'thing', with a limited number of options,

The options are finite, sure. But they number in the millions. Raise this to the power of the number of formations, and then add in the possible actions of the other side (bearing in mind sharply limited intelligence, whereas in Chess and Go intelligence is perfect).

quote:

..what toaw does now, is decide btn by btn, refering only to the pre-set objective track for that btn's formation,

This is more or less the problem. But the solution to this isn't a brute-force approach like you use for a Chess computer. You'd want a learning AI which tries to be more like a human player and less like an adding machine.

The idea I had would be for the player to design objective tracks independent of the formations. The AI would then judge how it was performing on each of these objective tracks relative to one another, and assign its formation to these tracks depending on the situation. So if the situation on one track is changing rapidly, that gets extra formations assigned to it.

IMO, the only solution if one wants to improve the Ai is to completely redo the OOB system of TOAW. Once you have a 'clear' OOB, ie this division is part of this corps, which is part of this army, which is part of that Army Group which is commanded by the Supreme Command, can you actually develope an AI that thinks in the grand scale.

Furthemore, the Supreme Command needs to 'know' how to give objectives(instead of designer implementing those objective tracks), then the army groups, corps, divisions, and battalions need to know what they can do and what they should do.

The key is information and how to use that information, and it should flow both ways. For example the 1st corps gives the 1st division the task of capturing the city of Kharkov. 1st division 'analyses' the situation and sends back a message that this is impossible without 2nd and 3rd division securing flanks. 1st corps then analyses the situation and decides if it can allow 2nd and 3rd division to do that. If not it will either cancel the attack or ask for more reinforcement...or if the situation is desperate it will force the order...and so on.

quote:

ORIGINAL: ralphtrick

quote:

ORIGINAL: berto
Where it makes sense, and where it is implementable in future versions of TOAW, please give the AI more time to "think" according to our, the players', choosing.

I'll do that, but the problem is fundamentally different from chess. In chess, the computer wins by figuring out moves 8+ moves ahead. In a game like TOAW, it isn't even possible to look at the next move in that kind of detail because the number of potential end positions in a typical turn is huge.

There are 2 options for a non-cheating Elmer.
1) Do like a human does, and try to remember where you've seen units, rely on your memory of past times playing the game and starting positions, and make some sound decisions about what your options are and what's likely to work, guided by hints from the designer.

2) It's possible to play Elmer with both sides so that he can estimate what you might have done, unfortunately because he play by the same rules you do, after the first couple of moves, that becomes very, very, very difficult.

#1 is much more practical, and is probably where I'm headed with Elmer, along with giving the designer greater control over Elmer's behavior. There are also things called influence maps that can help to analyze maps. I'm not sure what part of that will be in the initial release and what parts will come later.

Ralph

quote:

ORIGINAL: a white rabbit

..never the less, i can't see why it can't do objective-track comparisons and act accordingly, we've got the RAM to do the comparing,

quote:

..and no, you're still not seeing it right, the number of options open to a corps are very limited,

quote:

ORIGINAL: golden delicious


From a human point of view, yes. It can go here or here or here. But from a computer point of view it can put this unit in 145,100 and the other in 145,101, or they could swap round, or be in any of a thousand other combinations of hexes. All for one formation operating in a very small area. Then raise to the power of the number of formations.

Then add that you can't know what the result of any of these moves will be because your information will never be even close to being complete (even if you had perfect intelligence, combat is still random). So this brute force approach will never work no matter how much RAM you apply to the problem.

Not that formations are the problem. The PO can handle individual formations already.

quote:

ORIGINAL: a white rabbit
..and what's the fascination of the phrase raise to the power of ? 20 formations or units or whatever is multiply by 20, not to the 20th power, which is a much smaller number, one that he can now handle ..

quote:

ORIGINAL: golden delicious
How do you do this, though? How does the computer know what is important and what isn't? This is an incredibly human decision, which is why it is made by the designer
at the moment.

quote:


This is how real armies work- but I think it's an inefficient and roundabout way for the computer to work. After all, I Corps, 1st division and everyone else are in fact all controlled by one staff. Further, the situation is much more complicated that deciding whether or not to attack one point and whether or not to assign support to that attack.

quote:


So to take your basic concept, I Corps decides on a list of possible actions and sets them priority ratings. It then apportions troops to these tasks according to those ratings. Then it records the highest priority rating which could not be achieved due to lack of troops, and the lowest priority rating which was achieved. GHQ then reassigns troops from areas which were able to do all their priorities to areas which have lots to do.

All of this would be very difficult to make the computer do. The virtue of my proposition is that it fits the existing model where most of the problems have already been solved. It could also apply at least partially to existing scenarios as well as new ones.

quote:

ORIGINAL: a white rabbit
..and what's the fascination of the phrase raise to the power of ? 20 formations or units or whatever is multiply by 20, not to the 20th power, which is a much smaller number, one that he can now handle ..

quote:

ORIGINAL: Karri
I disagree, it's simple mathematichs. It's not difficult for computer to do. What is difficult is for the humans to program such a system.

quote:

ORIGINAL: Karri
Well here you are wrong, why not simply make many AI's? Why does the basic grunt need to be controlled by the supreme command?

quote:

We know what historically happened when a certain fuehrer though he was smarter than his generals and when he tried to order every single battalion personally.

quote:

I disagree, it's simple mathematichs. It's not difficult for computer to do. What is difficult is for the humans to program such a system.

quote:

ORIGINAL: golden delicious
Because the aim of designing AI is to produce a strong opponent- not one who acts like a human.

quote:


Hitler did not have an intel pentium processor and a gigabyte of RAM.

quote:


Semantics.

quote:

ORIGINAL: ralphtrick
So is cold fusion, we just need to figure out a couple of those pesky engineering details.

quote:


People are getting a lot better about programming AI. That's especially true now that PCs have enough power that they can devote some of that power into the AIs. Up until recently, especially for the real-time games, they were able to devote very little horsepower into the AI.

quote:


The big advantage to any hierarchical AI, which mimics the behavior of a regular army, is that it allows you to control the complexity of the picture that you need to look at at any one time. The General Staff AI needs to determine whether to attack or retreat, whether to head for the left or the right flank, and broad strategies like that. It can do this based upon a 3 x 3 or 4 x 4 representation of the map. It will also help the AI to respond realistically to feints, etc.

quote:


The main reason that you want objective tracks right now, is that determining the plan is an extremely complicated problem. A human can take a look at a map, read the briefing, and from this formulate a plan. While the AI can analyze the map, it can't read the briefing. You'd also need to have the AI understand events triggered by hex occupation, future reinforcements, etc..

quote:

ORIGINAL: Karri

On the contrary, we want one that can 'read' the game like a human. Not an AI that use this and that cheat to be 'strong', like in so many games so far.

quote:

Wow, you completely trashed my point. Your argument is pure gold.

quote:

Well yeah. But why does the 'general staff' still need to guide every single unit?

quote:

That's not even the problem, the problem is that the objective track concentrates on capturing the map, which causes a problem since the objective should be to destroy the enemy forces, which the objective tracks cause to be a secondary objective.

quote:

ORIGINAL: golden delicious
It doesn't have to cheat- it has numerous advantages anyway, because it knows the code in a way the player never can.

quote:


This whole approach to computers and robotics infuriates me. Why has a Japanese company made a walking robot? Robots are rubbish at walking- and tracks are better anyway. A robot will never walk as much like a human as a human does, and an AI will never be as good at being human as a human is. So quit trying- it's not necessary.

quote:


Well, to be fair, you didn't really have a point in the first place. It wasn't too hard.

quote:


It doesn't have to, but this is one of the advantages of AI. A human brain can never properly understand how one unit relates to the action of the other 1500. An AI can.

quote:


That's a good point. I suppose you should have the option just not to define any objective tracks at all. Then the AI locates the enemy centre of gravity and tries to neutralise it one way or another. It would probably be pretty good at doing the former through statistical analysis.

quote:

ORIGINAL: Karri

It doesn't know the code, at all. THAT would really require AI we've never seen.

quote:

...or you didn't undrestand it,

quote:

Nonsense. As it is right now, the AI is doing far worse job undrestanding the relations between the units. You encircle a division and the AI is happily sitting that other formation where it is, because that's it's objective.

quote:

The basic 'core' of actions always needs to be laid down by the designer. I still think the first step is to create a real 'OOB', instead of the system we have now, where each fromation is a separate entity.

quote:

ORIGINAL: golden delicious
I mean it can calculate odds without having to use the attack planner- because it is the attack planner. It can- in theory- mentally put all those units together and estimate the outcome. Humans can try to do this, but they're not as good at it.

quote:


I understood what you meant (but thanks for assuming I'm an idiot). You were picking a petty hole in my wording because you disagreed with other parts of my argument.

quote:


You don't seem to see what my point is. The above is the behaviour of the current AI. We're discussing alternatives. In the basic objective-based alternative I proposed, if the player suddenly leapt up one objective track (as would happen in a flanking or encircling action), the PO would react by directing more formations to that track.

quote:


I don't see this need. Since formations should be able to switch between higher HQs anyway, why tie them to those HQs for the purpose of the AI? It may be that it's best for the AI to chop the front up into sectors and assign its formations between them, but I don't see why this system should be based arround a rigid command structure.

quote:

ORIGINAL: Karri


In order to calculate the odds it has to cheat(see the enemy unit information),

quote:

yeah...which was a point I made aswell, so I guess neither of us see each others arguments...however I do not view this as the best possible AI, or even one that can offer much more 'resistance' than the current one.

quote:

Erm, because as it is formations have NO higher HQs. I don't really get what you are proposing, that there is a command structure or that there should be none?

quote:

Why would a proper OOB tie down the units to certain HQ? The point of a proper command system is just the opposite...as it is now, they are tied down, and have no higher HQs.

quote:

ORIGINAL: golden delicious

Is it a cheat when you use the attack planner?

The AI can make an estimate based on that same calculation (plus some bits which it currently ignores), only it doesn't have to move the units into place first.

quote:


I agree it's not the best possible- but I'm concerned with what is achievable, as there are numerous things that need doing and only very limited resources.

quote:


That there doesn't need to be. The reason for a command structure in a real army is that one human officer can only handle a certain number of subordinates. That's not the case for an AI.

quote:


They're not tied down. Take FitE. You can move a division across the map and it will perform exactly the same way with no other actions. If you design a force structure, you also have to provide a mechanism to allow units to transfer.

quote:

ORIGINAL: Karri

I ment that in order to calculate the attack it needs to see what kind of equipment the enemy unit has. Unless it can do that, it won't do any better than a human player. In fact, a human player can do far better because he knows the 'tricks' of the combat system. Ant unit attacks etc.

quote:

Well yes, very true. Since I don't play against AI I prefer the other changes take priority.

quote:

I'm not quite sure about that. It can handle a million units for sure, but it can't handle them well.

quote:

No, not actually. Supply, artillery support, etc. is reduced if the unit belongs to another formation.