Future AI

**John_Longarrow** · 07-13-2007, 02:45

Ramela,

One of the other big differences is that, as M2TW is a multi-faction game the AI, much like the human, can only work a few turns in advance for real moves. What he human does that the AI doesn't is to set goals and evaluate how relevant those goals are to the current situation.

As an example, as Milan several of my goals are;
1) Take Venice
2) Buy or Take Balogna
3) Take Florence

The order of those goals being achieved is dependant on what the AI does, how it moves its troops, and the results of several multi-outcome encounters. If I can't buy Balogna off the Germans, I'll have to change my strategy and take it later.

I would not be surprised if the first five turns of a M2TW game have more possible outcomes than an entire game of chess.

This means that the concept of setting goals and working to resolve those goals often is much more efficient for a human (or computer) than working out possible moves (the chess "Looking ahead"). As a result, I'm sure that by working out multiple AI scripts and having each faction leader adhearing to one of these we should have a net effect similar to a human player. Of course since the AI can all of a sudden decide on a very different style of play mid game, that can be a real challenge for a human to keep up with.

**phonicsmonkey** · 07-13-2007, 05:33

well, technically the number of possible moves in a game of chess is believed to be infinite

which means that no program or computer, however powerful, could simply calculate its way to victory

in fact most good chess playing programs have the benefit of centuries of human analysis of the game of chess built in, with standard openings and past grandmaster games on which to base their quantitative analysis of the "strength" of a move. these provide a guiding framework which help the program to select which candidate moves to spend its time considering..

this should put into context the stage of development of game playing AIs that play other games like M2TW.

chess AIs have a head start of several centuries of human knowledge. it's unlikely M2TW (or any similar computer game) will be played for as long and be studied so deeply and comprehensively by so many highly intelligent and talented human players (we at the Guild can only do so much!)

this is not to say that it's impossible to create a challenging M2TW AI, but it should go some way to showing that the level of AI achievement is actually quite high when you consider the starting point - weeks or months of testing, rather than centuries, have gone into creating a set of rules to guide this AI in playing M2TW.

**ramela** · 07-13-2007, 07:53

@ John, I actually referred to the nuber of factions and "moves" as a problem for making any kind of "situational awareness" for the AI, as those things make M2:TW infinetly more complex than chess, which has been studied.

Personally, I dislike forcing pre destined goals on AIs because take away from the "every game is different" aspect that I like in games, but I agree that making the AI more goal oriented approach is neccesary. Plans carried out only half way are worse than no plan at all. In Didz's example the AI should have (propably) accelerated his time table for assaulting Sofia and Bran, hopefully sack them and then decided whether to go for defence of the homeland or continue the assault to Constantinopole and/or Thessalonica.

I think behavioral modification, possibly even copying the humans behavior and not just reacting to it, could be useful here.

@ phonicsmonkey, I think the number of moves in chess are not infinite if you use the "After either player has made 50 moves, during which no pieces have been taken and no pawn has moved, the game is considered a draw" -rule, since due to pawns moving only forward, sooner or later you will run out of pawn moves and eventually you will run out of pieces to take.

For more useless thought play on chess, read the below. As always, if I messed up, feel free to prove it and enlighten me.

Spoiler Alert, click show to read:

I also left out opening databases, end game databases and search tree pruning from my post on purpose. for simplicity's sake.

I agree that for the reasons you state M2:TW's AI will never see the kind of development that goes into machines like Deep Fritz, which is why I think highly of any computer game that has an AI at all. :)

Except, technically M2:TW is not an AI, since it couldn't pass the "Turing Test" against anyone who has played for a significant amount of time.

**Nebuchadnezzar** · 07-13-2007, 08:46

Originally Posted by ramela

M2:TW infinetly more complex than chess, which has been studied.

Studied by who? and what are their qualifications?

I think the suggestion that M2TW is infinitely more complex than chess needs to put into a more realistic perpective. Also are we talking about the battle map or campaign map?

If its the battle map then I don't think so. Every group of units is essentially acting as a whole and not individually. It is only an impression that you are controlling 1000's of units rather than 20 or less. Furthermore, each unit moves one tile at a time. There are no units that magically transform to the opposite end of the map in an instant such as they can in chess (eg bishop, queen etc) unless its a bug. Finally most units are more or less identical with only minor variations. Cavalry are cavalry, infantry are infantry and archers are archers. Just how many different variations can exist. Not many.

The campaign map is of course much more complex but not infinately more complex by any means. The available pieces on the campaign map do not all interact with each other but rather with only a select few if lucky, diplomats for example only interact with other diplomats, generals, princesses or settlements and they only open a diplomacy screen with a dozen or so options. Also it is essentially a two player game with player Vs AI. Any AI Vs AI interaction is more coincidental or chance or for player amusement rather than any greater plan or strategy for the AI. Too watch AI Vs AI is one of the saddest things in TW games!
But its probably the huge size of the game board and the no. of tiles that create the impression of complexity but consider that barely 2-3% is used at any one time. Again there are no units that instantaneously transform to the other end of the campaign map as they do in chess.

So a finite number of pieces present, each able to be moved only a set number of tiles per turn and each piece having a very limited response and interaction overall. I think any well trained chess player would laugh at the suggestion that M2TW is infinitely more complex.

**ramela** · 07-13-2007, 09:26

I meant that chess has been studied. By many very qualified people, I would Imagine.

EDIT: In terms of "pieces" and possible "moves" M2:TW is much more complex. (How many agents, military units and so on do you have on your firt turn as the Milanese? What about turn 10? How many different tiles can you leave each of them on at the end of turn?) I used the word "infinite" as short hand for "too great a difference to understand and describe".

Also, I am talking about computational complexity, specifically in terms of arithmetic computation that a computer uses. Humans are a lot more suited to analysing large masses of data like the position of several armies with 1-20 units around several settlements. Computers are not very suited for evaluating such data. A Chess position is still very complex and thus a human has the advantage in that respect too, but a chess position is sufficently simple that a computers raw number crunching power comes into play here.

Still, it takes a super computer specifically engineered for calculating chess positions to beat the human world champion. A proof of how complex chess is.

Luckily for CA, since M2:TW revolves around attacking armies and settlements, focusing on those simplifies AI design a lot.

**The Badger** · 07-13-2007, 09:56

chess:
all valid points - and all points i think ai folks have discussed b4 - about other wargames - that chance and variance add EVEN more factors and that even seemingly simple chess has so many variations...
even the 'risk' map of stw/mtw is difficult to picture ai scripts for once you try to imagine all the branches - this addresses the 'have a plan' issue the other fellow brought up - full circle; back to the point that chess ai's use pre-defined 'base' strategies (which i did not know)

ai vs human players -
the warcraft fellow made 2 good points; one is that humans are not always available or desirable; and sometimes no better
but i also feel that alot of the ai work ignores that the greatest asset mmo's HAVE is the non-artificial intelligence which i s NOt in short supply...

i have always hoped the campaign game would evolve to where a triggered battle might be POSTED online for an opponent - and ranked players would qualify for certain rank 'bad guy' generals. Its probably DOABLE... but would it interest anyone BUT me?

ANCIENT ART OF WAR
was a pc game for DOs that had several different computer opponents each with a somewhat different ai script - they actually acted differently; not just more or less difficult (if i recall correctly).

MTW - im not sure about m2tw - had broad strategies defined for the powers - i was never sure how really different they were.

But i think further developing THAT would assist with the predicatbility the one player spoke of.

**Didz** · 07-13-2007, 10:16

Originally Posted by The Badger

i have always hoped the campaign game would evolve to where a triggered battle might be POSTED online for an opponent - and ranked players would qualify for certain rank 'bad guy' generals. Its probably DOABLE... but would it interest anyone BUT me?

The lack of a multi-player campaign option has always been my biggest dissappointment with the TW series and probably the reason it has never been as massive as it could have been in the wargame community.

**Shao** · 07-13-2007, 22:08

The Badger, I do remember The Ancient Art of War! What a great game it was... I may even have it somewhere deep in an archive on my HD... :-)

The opponents there were supposed to act differently, yes. Their descriptions gave hints about their styles of warfare. I don't remember if they really followed the patterns though.

**Didz** · 07-13-2007, 09:51

I don't think anyone was suggesting that MTW2 was a simple as chess, the point I was making was that because Chess programes have developed a level of situational awareness they are able to make better decisions.

The real issue with MTW2 is that its situational awareness is practically zero, as per my example of the attack on Budapest. Therefore, its ability to plan appropraite actions and counter actions is also practically zero. Everything that happens to the AI in MTW2 is a virtual surprise and it has no way of planning its long term goals as suggested by John.

By comparison I was given a pretty good demonstation of situational awareness and goal planning by the AI routine that drives the Combat Mission engine only last night.

Just for a laugh, I decided to set up a custom battle which pitted my two King Tiger tanks against a bunch of low quality US infantry. My first mistake being to allow the AI to choose its own mix of infantry units. Needless to say it chose a lot of bazooka teams, proving that it had considered the opposition and planned the best army composition it could within the limitations I had imposed upon it. Something which MTW2 doesn't do for a start.

When the battle started my King Tigers positioned themselves hull down on the ridge guarding the objective, giving them a commanding view of the open ground to their front and flanks beyond which were the woodlands from which the enemy were expected to emerge.

Six turns went by and not a target appeared, until with a sudden 'whoosh' from a copse just below the slope a bazzoka round bounced of the front armour of the nearest Tiger. That was a bit of a shock and both tanks pounded the small copse to matchwood with 88mm HE and machine gun fire until the US Anti-tank team tried to run and were finally cut down.

After that there were a lot of sporadic sightings of US infantry running back and forth amongst the trees and the tanks were kept busy engaging multiple targets along the edges of the woods. By now it was turn 22 and the worst seemed to be over, I was thinking of advancing down the slope and mopping up the few remaining infantry contacts.

Suddenly, there was a dull plop and a mortar round dropped onto the ridge next to one of my tanks. I shrugged, a small calibre mortar round was no threat to my tanks armour, I figured it was just the AI getting desperate, but as the number of rounds increased I realised that they were not trying to damage my tanks, the rounds were 'smoke', they were trying to blind me.

As the ridge gradually became enshrouded in smoke, my tanks found themselves unable to see anything, and assuming that this was a prelude to some sort of massed charge by grenade weilding infantrymen I decided to reverse further down the slope and get clear of it so that my tanks could see what they were firing at.

As the Tiger on the left reversed down the slope there was a sudden 'Whoosh' sound to its rear and bazooka team which had inflitrated behind my position and located itself in a small wood put a round straight through the rear engine compartment. They had actually been sitting there waiting for the Tigers to reverse down the slope into them.

Fortunately, my other tiger had chosen to reverse along the road and had so had avoided the trap, but the enemy infantry had now moved forward and taken the objective and once the smoke disspated I had no chioce but to advance and re-occupy the ridge. As the remaining tiger advanced it caught yet another US bazooka team sneaking across open ground towards a farmhouse on the left of the road, presumably to set-up another ambush, and managed to elminate them. It was then forced to button-up by heavy machine gun fire from the ridge and found itself under attack from multiple anti-tank teams of several different angles. Inevitably, a lucky hit penetrated the side turret armour and the battle was over.

By comparison with MTW2 this was a challenging battle, and the AI showed not only good situational awareness, in its army compostion but was obviously working to a good tactical plan that made the most of its limited assets. I suspect it was using the 'universal knowledge' cheat but thats one of the lesser issues with AI routines. I'd certainly be much happier if the AI for MTW2 gave a similar level of challenge.