I downloaded the new demo and have been thoroughly enjoying playing the game. I noticed however that AI difficulty level can only be set when the scenarios radio button is checked in the game type dialog box. The impression I had from the release notes and other messages on this forum was that the AI difficulty level applied to all game types but this impression appears to be mistaken. Can anyone clarify whether this is indeed the case.

So far I'm very impressed with the game, impressed enough that I've ordered the full version. Congratulations to the developers on creating such a great game and on being so responsive and involved with the user community!

Thank you very much for such words, which are like balsam. The difficulty levels for World Wars is quite a tough question. The problem is that on Emereal EASY level of AI was harder than HARD just because often A.I.'s random unit selection happens to be more effective than strictly calculated variants. So this demo has difficulty levels for scenarios only.

But I only just downloaded the demo the other day! Arghh! So what difficulty does the AI use now? While its unit choices were better on easy (missile trucks ), its placement of units was certainly not.

Basically, where do we stand with this now? Has the AI been tweaked for that map alone, or are we stuck with the hard that isn't really hard for that map?

So you're saying that AI difficulty levels are based on the unit selection strategy rather than on different levels of tactical sophistication and it just so happens that on the demo world map (Emerald) that a bad unit selection strategy is actually sometimes a brilliant strategy?!

Easy AI has a good chance to buy a rocket Laucher near the border and it's considered EXTREEMLY ineffective by expert players. However starter players need time to learn how to cope with this situation - that's why it seems more effective on EASY.

The depth of tactical calculations also changes with the AI level. On easy level it uses the first move combination it finds without trying to change it to be better. But that makes little difference for noobs because all AI units don't stay and wait, but fire at players units.

So, if difficulty selection is gone now (can someone run through why this was done again?) what level will it operate on?

As I said, its unit selection was better on easy (sort of... too many rocket units too early ) it was obvious that its tactical behaviour was much worse Reassuring to see the difference in action

I don't think removing difficulty levels at World War level was a good move though.

Anyway, my eleven year old brother is currently at 150% against the AI on easy, so it's not that bad

/shrug
i think playing worldwar against AI is kinda...useless thing because AI cant do anything meaningful in strategy scale and will inevitably lose to even an average player..so does it matter is that AI "hard" or not - its easy anyway...some sort of singleplaying challange are hard scenarios if you like that sort of thing...

Then you will not mind, in your indifference, if I continue to care about this issue? Cheers.

The fatal flaw in your post, which you make in all your posts, is that you assume the AI is a method of training before coming online. For many people this is not true, the AI is the beginning and the end of their gaming, and for others, the AI makes for a completely seperate game which has nothing (and they like it this way) to do with multiplayer.

Of course, more importantly, single player gamers currently, and always will, massively outnumber the online gamers. AI development is a rather important aspect of this game.

Quitch is absolutly correct. The interesting bit is that of course there are more MP fans on the boards and forums than there are SP fans, mostly because MPers are already online, while not all SPers bother to go online (much or at all...).

I heavilly lean toward SP myself, though I appreciate the added challenges of MP, however, time is an issue for me, and that often means that I cannot get to the game for a meaningful amount of time (to make MP turns) for periods of days, thus I am often loathe to begin MP endevors out of consideration for my opponents, who want a faster pace. For me its a *bonus* to have MP not SP, SP should always take precedence as it is the best way to increase the appeal of the game to more people, and thusly to bring more people into the MP community. Dev teams that ignore SP (mostly AI issues...) are doing the entire community a serious misdeed (not that the devs here are doing this ).

Quitch is absolutely right. We, the developers, do clearly realize that the majority of the players will never go online just because they got used to playing singleplayer and they are quite comfortable for them. And MA's A.I. must be interesting and challenging for them.

That's why we have two DEDICATED developers, both Computer Science "With-Honors" Graduates, who are full time developing and upgrading the A.I.

Is there some learning in the AI strategy ? As I'm working in an AI lab, this subject interests me a lot
The main problem is the majority of the games is that Ai is fixed and cannot evolve. In this sense, once you got his strategy, its really easy to beat it. I know that learning schemes are not easy to implement, but I would be rather interested in technical discussions about this subject

Ahhhhh, one of the subjects I love! I took every AI class I could in college, and even did a little primary research. Though I don't have much work experience in the field (5 months in a game company, and even that was limited), I soak up everything I can about AI. I'd be highly interested to discuss this as well!

I seriously doubt MA uses any learning, as it already requires a huge amount of resources I'm sure for what it already has to calculate (my guess anyways). It does seem to stick to some basic strategies, but doesn't seem to be too stuck on one pattern, as it can react fairly well to how you manuver your forces. Seems like a descent balance between predefined strategies and reacting to the current situation. It protects its weaker units, invades when it has a force gathered, and chooses where to place its purchased units very intelligently.

The fact that the AI's units work well together leads me to believe it isn't entirely reactionary. But then again, sometimes it will break its basic rules to go after a good target... (like move LAVs back from the front lines so mortars can come up and hit a hurt heavy bot)

Oh, THAT'S WHY!

Damn! And I've been thinking all that time why, why why?

That's why, you stupid Victor... that's why, man....

I suspect that the AI exists as two entities, tactical and strategical, and that these two talk to each other to work out the turn. The tactical AI would make decisions on how to use the untis available to it. I suspect it first tries to block the border, also tries to guard vunerable units, and then calculates how it can do this while inflicting maximum damage.

The strategic AI decides how to buy units, where it needs them, where it should move its units to, which territory to reveal, etc.

I would strongly doubt the AI learns. AI learning is very complex, CPU and time intensive, and is not really a good "value for money" idea. It simply isn't advanced enough.

AI learning, on a far more basic level, could be done though if the game took note of what players did in multiplayer. Say for example two players were on Emerald. If one of those bought only LAVs and Rocket Launchers, while the other did LAVs and Tanks, and then the first player won, the game would note this, inform a central AI server, which in turn would update the AI to weight Rockets Launchers and LAVs higher on this map (and the amount it changed the weight by would depend highly on the ranking of these players... privates would be virtually ignored, while a Marshall would have a real effect. It could even file a replay along with the result this replay had, so that if the AI designers though "Why does the AI suddenly love RLs on this map?" they could see why). This server would then update the AI of any client which connected to it (via an Update AI button).

Of course, this would take a little work, and the way in which the player used the unit may differ from the way in which the AI uses it... but then that's why the AI designers should be watching replays of the best players in action, and comparing that with how the AI uses the same units... perhaps even going so far as to setup situations identical to those in the recordings and seeing how the AI would react. If it reacts in the same way as a Marshall, then their job is done

This is all just a rough outline of course, something I once brought up for Total Annihilation (should have written an article on it ).

Very intriguing idea! An AI that would watch every game to learn from humans... With the fact that the Massive Assault server sees every single game that is being played it would give it plenty of trials to look at. And the more data a learning AI has access to, the better.

Couple issues that would need to be overcome though: Say two weak players are playing each other? The AI would more likely learn bad habits than good habits from these match ups, as newer players haven't quite gotten the handle on what units to use when. I would say the AI would only consider battles between two higher ranked players.

Why not also let people apply it to single player as well? In single player wars, it could first determine who was the winner and by how much, and only consider battles in which the player handily defeated the AI. The AI server could then go and try to find out what let to its defeat, weighing it based on how bad its setup was in relation to the human opponent. Of course, it should be optional whether the player will use these games to help the overall AI, so an internet connection isn't required, but used if available.

I also like the update button idea... some kind of central AI server is just processing games and trying to learn from them. That would have to be one heckofa extensible AI structure though, if it was going to learn more than just unit weighting. Unit weighting I don't believe would be that helpful unless the AI could also determine the full circumstances of why a particular unit was most useful in that situation.

As I say, a private (or whatever the lowest online rank is) would be virtually ignore, while Marshalls could swing things over the course of a few games.

I am trying to keep it simple (i.e. practical). Unit weightings are easily done, and the tactical AI in this game is already excellent, leading me to think that as long as the AI was given the right units, it would make good use of them. Replays can aid the designers in improving the AI on this front anyway

Teaching the AI on a strategic level is something else, and far harder to do. There are so many factors involved in each strategic decision, that you'd only be able to get the AI to check a few, and would probably end up giving it a slanted view of the world (say a player reveals a secret ally and goes on to win... the AI looks at the situation and decides that it seems to be because there were enemy units of type X on the border... but really it was because the player had decided the game was in the bag and just wanted to win faster; had decided the game was lost and wanted to go out with a bang; wanted a heavy bot on an entirely different front and was bringing a transport back for it)

Too many variables IMO. My method is little more than an AI "second guessing" method. If you want to hit more complex areas, then the AI needs to learn in other ways, say, fighting itself in game after game (switch off the graphics and let them rip) revealing secret allies at different times in different situations, recording various variables and deciding which ones worked out best, etc.

At best it could monitor things like "When he revealed secret ally X, the country control situation was Y" then record this for every game, and try to find a "best match" scenario when playing: "Country X is mine... according to my template database I should reveal it when the country situation is like Y, Z, A, B or C". In fact, you'd probably be better hand-writing situations for the AI, like "when the enemy invades secret territory with > $X troops, wait until value falls below $Y within a Z hex radius of the territory before revealing. And so on.

What I'm really saying though is that the weighting method is simple, and while there are hurdles that you would need to be careful of, it can be done in a reasonable amount of time. I'm not so sure that such a limited, and simple learning procedure, would work in the case of the strategic level due to the extra levels of complexity found there.

However, think of enough variables to monitor, and maybe it can be done.

As this is a long post, please excuse my bad english

2 Vic :

2 Maelstrom and Quitch : sorry guys to answering so late but I was quite busy these last time. It is really interesting to read your posts and I think you've nice ideas Futhermore, it's always great to discuss on such a subject with passionated guys like you. Quitch, are you also working in the AI field. If it is the case, I would be interested to know in which field more precisely (personnally, I'm working in artificial neural networks and in dynamical system and chaos theory).

Concerning the debate about learning in games like MA, here are some of my ideas on the subject :
First, I think that a turn by turn game is the ideal terrain to elaborate some sophisticated AI because there are much less computational resources constraints than some others type of game (as for exemple a real time strategy game).

Concerning the AI probel in itself I'll consider a quite general framework : after perceiving some informations on the state of the game, a “reflexion” of the AI must lead to an action which will change the state of the game.
Of course, the main problem is to choose the best action between all the possibilities. In theory, one can really choose the best action because each possible state lead only to a limited number of others possible state and as there are final states it is possible to evaluate the true value of each state and so one can really choose the best. In practice, it is another story. The number of possible states is so huge that it is completely impossible to gain access to the whole tree of all the possible states. Because of this problem, one cannot have access to the real value of each state. The only thing that can be done is to try to predict this value. For that, one can imagine a lot of measures (or combination of them) to evaluate the value of each state (e.g. the number of enemy units points destroyed, the number of units exposed to enemy fire, …).
Now, I think that human players use also some heuristic evaluation of the different possibilities they have when playing. It is clear that human player have also learning capabilities which allows them to refine their initial skills by experimenting a lot of games (just think to our imitation capability : when we see a successful opponent strategy, we can learn it to replace it in future similar situations). Concerning this learning ability, I think some of the basics could be incorporated in an AI motor. For example, one could imagine that the system learn new way to perceive information and new way to valuate this information. Concerning the perception stuff, I think that all the available information concerning a given state have not the same value : some information have very little value (as for example the color of the units ) and some others have much more (as for example the relative disposition of the enemy countries and the allied countries). Some of these informations are high level information in the sense that they are combinations of more basics informations. In this framework, one can imagine a system which learn what are the valuable information and what are the information to discard.
Concerning the way to give value to a state, there exists also such a wide range of possible measures that it seems interesting to give to a system the ability to adapt the way it valuate a state by learning new way of measuring value or by refining his initial schemes. All these learning strategies could be implemented by various mean (e.g. reinforcement learning (that is, feedback mechanism allow the system to reinforce certain behaviour and to discard others), heuristic search by means of heuristic algorithms as genetic algorithms, …).

These were all basics and general ideas and of course, putting them in pratice can reveal really difficult.
Concerning MA itself, I don't know how the AI motor works, but I'll be rather interested to know more on it. I think that the AI motor in this game is really impressive (thats the best AI I've ever played ). Just one little remark about the fact that his most weakness (to my opinion) lies in the fact that it doesn't use transports at all. As mobility and timing are key concepts of the game, I think that maybe some amelioration could be done in this area. Nevertheless, congratulations to the developper to have created such an impressive AI motor

well actually...AI uses transports...
dont glare at me like this - YES it does!
for example download the "NoMercy" scenario
http://massiveassault.wargamer.com/nomercy.html
or look at my replay attached below.
the AI had two transport and used them...for a very bad result. - i was able to destroy one of them with a rocket launcher inside...
It seems that AI can use transports IF they are pre-deployed by scenario designer.
But AI not knows where and then he should purchase transports.And AI is unable to predict the attacks from players trasports and of course AI is unable to predict the usage of "transport chains"....

Oh, SHIT!
I didn't know it was SO bad...

Maybe that info will make me surrender in our game on New Paradize...

2 Pitor: Thank you alot for your warm words.
Exactly at this moment we are hardly working
on using trasports by AI.
So, be ready to fight