The AIs

[daviddunham]

Battle of the Bulge actually features 8 different computer opponents: 2 per faction per scenario (the “Von Rundstedt” AI has the same name in “Race to the Meuse” and “Battle of the Bulge,” but is tuned differently since they require very different play).

Mostly we designed them to give different play (at a reasonable level of challenge). But we didn’t label them as “Hard” or “Easy” opponents, partly because we didn’t set out to make them that way, and partly because we don’t have a specific answer.

So, is there an AI you’ve found especially easy to beat or who gives you a serious challenge?

[jeffd]

I'd just like to add to this, as one of the two people most responsible for the AIs: we will be watching this thread and taking notes. So if there's something that the AI is doing well or poorly, please let us know so we can make Bulge and later games better.

Cheers!

[daviddunham]

If you haven’t seen Miguel’s article on what is behind the 8 AIs, check it out!

[s_pixelgeek]

If you haven’t seen Miguel’s article on what is behind the 8 AIs, check it out!

Interesting and quite technical

[Yojimbo252]

At the request of David I'll transfer a post I made previously which touches on the subject of the AI with a slightly different spin...

I applaud the amount of effort you guys are putting into strengthening the AI.

However just to put a slightly different perspective on things (and I appreciate this may be just my personal opinion) but I would suggest single player for a large number of players will tend to serve as simply an extention to the tutorial to learn the basics.  Once a player feels they have a reasonable grasp of the mechanics they'll likely move on to multiplayer and rarely look back, as multiplayer is where they expect the real challenge to lie.  

I don't think any experienced gamer reasonably expects the AI to be able to rival a good human player as iOS game developers don't have the resources and time available to create a 'Deep Blue' for their game.

So you may well find that upon release whilst a portion of players will demand the AI to be strengthened (which is usually the case with any new release) there's likely to be a greater number that will request improvements to the multiplayer aspects like chat functionality, next game button, game/turn timer, and perhaps longer term improvements such as a player profile recording win/loss %, leader board, ELO rankings and rating based matchmaking.

The point I'm trying to make is that as I'm sure you're aware, AI improvements tend to work on diminishing marginal returns.  The stronger it gets the more time you have to invest for that next increase and each time you tweak it you never know what you might break.  The good news with multiplayer enhancements is that the effort involved to deliver a particular piece of functionality is easier to estimate and the finish line for that aspect is quantifiable.  Additionally any gains made will also enhance subsequent titles such as El Alamein whereas AI tweaks may not be so readily transferable.  So it may be worth bearing this in mind when formulating the development priorities upon release and going forward.

I accept that I may receive some backlash from people that are more interested in single player rather than multiplayer as it boils down to personal preference but irrespective of which you prefer, I'm sure most are aware that The Coding Monkeys have set a high standard with Carcassonne in terms of multiplayer functionality and it's one that BotB will likely be measured against.

[wmm77]

I have only played as the Axis so far and I have won against both AIs, Montgomery and Patton without much difficulty.

The AI really seems to stumble in one major area; it allows units to easily be cut off from supply. This has been the allies downfall in every game I have played.

When moving and attacking, the AI often leaves key units vulnerable to becoming out of supply. On three separate occasions, the Patton AI attacked or advanced into an area with 3 good units with no secured supply route making it easy to cut them off. Since they have moved, there is nothing more to fear from them this turn, and I simply ignore them and go around. Voila, they are cut off and OOS the next turn and unable to move. The AI does not seem to try to re-establish supply to those units. Likewise, I have had situations where my supply could have been cut off and it wasn't.

Does the AI consider supply when moving defensively or offensively? Does your AI take into account how much time is left in the day before dusk? Does it analyze possible enemy moves in response to its own or only the best possible position following its move?

I commend you for trying to establish the best AI possible. I bought the game primarily to play solitaire. I expect to get some on-line and pass-and-play in, but I think most plays will be against the AI. I would not mind if I had to wait longer for the AI to move if that would make a difference. Thank you for not taking the easy way out by fudging with the random number generation, slanting hit percentages, or inflating strengths for he AI. I wish you success in coming up with other AI opponents that offer more challenging play, or in tweaking the AIs you have to take into account where their shortcomings lie.

I think this is the best wargame I have ever seen for the iPad. I have played board wargames for many years and this is a breath of fresh air for the iPad. It is like some wonderfully awesome mating between board and hand held. The interface is great. The graphics are great. All the bells and whistles you put in it are great. I feel guilty that I did not fund it on Kickstarter. Rest assured, based on this gem, when you're ready for number 2, I will rush to fund it. Thanks or a great game.

[jrayls]

Hi -

Thanks for the terrific game. I am really enjoying it, and I'm playing my first online match now. I'm looking forward to seeing what you guys come up with in the future. If only we could get more "real" wargames on the iPad.

I've found the AI opponents to be particularly easy. I'm not really sure why. I've beaten all four on the full scenario without losing. I think someone mentioned it earlier, but the AI seems particularly prone to being cut off. I played (maybe way too much?) Unity of Command, and the cutting off supply tactic works quite well here as well.

As I'm typing, this internet match is turning out to be interesting.

Thanks for the terrific game. It's really a gem!

Josh

[blahblah3502]

My 2 first games were against the "aggressive/risk-taking" AIs, Dietrich and Patton.
I was surprised (I am not an experienced wargamer) to win fairly easily against both (winning both times before Christmas, I think). This may be due to these AIs being more prone to randomness due their risk-taking however? Maybe another time the dice rolls will go their way more.
I've since been playing against the "prudent/methodical" AIs, Monty and Von Rundstedt, and I'm struggling (have yet to win after 3 or 4 games or so with them )

all these games were Battle of the Bulge scenarios rather than the Race to the Meuse

[daviddunham]

All the AIs do consider supply. We hope to improve their consideration of it in a future update.

[Croix de guerre]

Hi AI guys!
This post will probably sound over critical, but it is only because I enjoy this game and wish the next ones in the series to be even better.

I rated the game four stars because currently I feel that the Bulge has value as a pass-and-play game only:
- the Game Center experience has serious pace issues - I have not been able to finish a game yet
- the solitaire experience as poor replay value: AI is not strong enough to keep being entertaining.

I have read the article about how you designed the AI. I understand that you implemented the AI using scripts and other techniques from first person shooters and deliberately avoided tree searching techniques from other fields - chess, go - that go back to the birth of computing.

My opinion is that you will never get an entertaining AI without a dose of planning - i.e. tree searching.

Here is an example, which happened just yesterday. I was playing the Axis. I managed to get a unit beyond the Meuse and trace supply to it before December 19th. The AI did the single worst move possible: it passed. I passed. End of the day. I won. I felt cheated of a proper victory.
Planning into the next move would have been sufficient to see that it was a bad move.

An additional benefit of tree searching is that it can adapt to different scenarios - even random scenarios -, unlike scripts.

The search space for Bulge doesn't seem too large: few units, few eligible moves. It seems larger but comparable to chess, much smaller than Go. Have you studied the Computer Go literature? Have you read about Monte Carlo Tree Searching? If yes then why didn't you go that way?

[daviddunham]

You can read about the AI at http://www.shenandoah-studio.com/aiutil ... haracater/

I'm not quite sure what you mean about Game Center pacing. Any remote turn-based game is going to have delays. Given you can have up to 30 games running, there’s a reasonable chance that you won’t have to wait too long. (Indeed, with only 3 or so games running I can find myself really barraged with turns if a playtester is devoting some time to the game.) As Axis you may have to wait for an Allied pass the first day or so, but it picks right up.

[Airshark]

I've been unable to get a GameCenter game finished yet either, though I admit I haven't tried that hard yet.

The game is excellent for immediate pick up and play feel, but the AIs are far too weak to challenge a serious gamer, particularly when playing the Allies. I played about a dozen games against each Allied AI, and won most of them right out of the gate; at this point, it's mostly a matter of very, very bad luck to lose now. So on I moved to playing the Allies.

Umm, this is where I think it needs some serious work. I've won each and every game I played as the Allies, even when I made an obvious rookie mistake, even when I had some atrocious luck. In one game, I managed to put all but one unit in the entire German OOB out of supply, and that one unit was a newly-arrived unit on the board edge. The AI will throw an entire SS panzer corps (three SS panzer divisions) against a single American infantry division on the flank, even if it leaves an immediate hole that allows that corps to be surrounded. After my first game, an easy win, I noticed that Dietrich didn't really even pay attention to being surrounded (I understand that's part of his strategy, but my take is that he'd just as soon be surrounded as not), so I started leaving open holes hoping the AI would dash in and let itself be surrounded. Worked almost every time. Then I took it to the next level - I started leaving weak units in spots that I hoped the AI would attack, win decisively, and use advance after combat to advance into a surrounded position. That pretty much worked every time. Runstedt's no different, really, you just have to play very slightly differently.

Playing the Germans is somewhat of an intellectual challenge, but the player should win most of the time. Playing the Allies is unfortunately a waste of time. If I had a suggestion to make, it'd be to pay a lot more attention to not getting surrounded. Or maybe change the rules so that breaking out through undefended zones is possible.

[Amaranthus]

Airshack, I agree that the AI struggles against serious wargamers (it hasn't beaten me yet), but the game is quite balanced for human opponents. Let me know if you want to try an MP game, I can play allies. We'll see then whether your Lehr can punch my line!

[s_pixelgeek]

After my first game, an easy win, I noticed that Dietrich didn't really even pay attention to being surrounded (I understand that's part of his strategy, but my take is that he'd just as soon be surrounded as not), so I started leaving open holes hoping the AI would dash in and let itself be surrounded. Worked almost every time.

This is my experience as well. So much so that I don't play Dietrich any more as its too easy to sucker him into making a bad move

[Miguel_TSS]

Thanks so so much for really taking to the Single Player experience we've crafted. It is an absolute delight hearing all your experiences and strategizing.

Some of you have run into and reported a bug I'm aware of in the Architecture. It is fixed in our next update, but if you run into scenarios where you believe that an Agent should have made a different move in, please keep sending them my way ( Tap-Tap-Press on Dials to submit a report ). I review all the games and love watching you play.

Now for a quick peek under the hood to let you know a few methods behind the madness in our decision to focus on PLAYER rather than WIN.

New Player Education
Bulge Abstracts Movement, Combat and Supply differently than most other games in the genre. We wanted new players to be able to see an example of how these can be used, but also encourage them to discover strategies that they could take to human opponents. "Easy, Medium, Hard" is not a strategy and has no place in modern gaming. By thinking "What would XXX type of human do in this situation?" We came up with an alternate set of goals other than win.

Problem Simplification
It may seem that I hate tree searches, that is not entirely accurate. There are too many factors that go into what is a "Good" Move. By breaking down the problem and answering into Archetypes, it became much easier abstract game data into the decision making processes.

Early( Poorly implemented ) Search algorithms can show very poorly, without good pruning they take too long to run. Without proper depth, good heuristics, or an easily definable win, they yield poor results. Having the AI take baby steps gave measurable results not only at sprint review, but each day of development. Heuristics are all loosely coupled so they could be tested, modified, or removed entirely allowing for rapid prototyping. We didn't have infinite time so we implemented something reasonable that lets us grow.

But it's TOO EASY!!
Nah, You're just too good ^.^ Our architecture makes it very natural to build making smarter, meaner Agents. Having created modular pieces to decision making, the programming team can work independently of Agent Designers. We love all the feedback, because that determines which direction this (and future) games take.

[Theodotus]

I find the AI to be about equal to the ability level my human opponents have demonstrated so far. So I'm content with using it for practice in preparation for multiplayer matches.

[Croix de guerre]

Hi Miguel! Thanks for sharing that!
I praise your willingness to answer negative posts constructively.

New Player Education
Bulge Abstracts Movement, Combat and Supply differently than most other games in the genre. We wanted new players to be able to see an example of how these can be used, but also encourage them to discover strategies that they could take to human opponents. "Easy, Medium, Hard" is not a strategy and has no place in modern gaming. By thinking "What would XXX type of human do in this situation?" We came up with an alternate set of goals other than win.

I am not sure to understand. Do you mean that the AI serves two different purposes: as a learning tool and as a computerised opponent?
Having the AI to learn the game is very valuable indeed.
On the other hand the AI is currently just to weak to be entertaining after a few games. It has a direct impact on the replay value of the game.
It is fine to have an AI display different personae as long as it is strong enough. It is not the case right now.

Problem Simplification
It may seem that I hate tree searches, that is not entirely accurate. There are too many factors that go into what is a "Good" Move. By breaking down the problem and answering into Archetypes, it became much easier abstract game data into the decision making processes.

When I play the game I find myself thinking one or two game-days in advance. I do not believe an AI could be any good without any planning at all.

You have built a very valuable domain-specific knowledge base with your scripts, rules, heuristics - I don't know how you call them. You already have done the most difficult part.

Now, you can leverage that effort by plugging this work into a Monte Carlo Tree Searching system. Let the MCTS search the tree by evaluating nodes by running simulations using your current AI. You don't even need an evaluation function!

To improve strength and response time, build an offline opening book, using the MCTS.

Finally, move the computation to the cloud. Increase the size of the opening book 1000-fold, or - better - keep refining the book live constantly. Most AI requests will hit the opening book and thus will be answered instantly. The misses will be carried on by the MCTS - a few seconds - and will participate in refining the book.

You can still have different personae by having a separate server for each persona.

[Miguel_TSS]

I am not sure to understand. Do you mean that the AI serves two different purposes: as a learning tool and as a computerised opponent?
Having the AI to learn the game is very valuable indeed.

Yes. The Agents use strategies that are understandable and can be imitated by new players. There is immediate payoff for a player to why an Agent makes a move. They are also abstractions of common strategies we've seen during development, so its a good way for players to experiment to see what works best for them.

On the other hand the AI is currently just to weak to be entertaining after a few games. It has a direct impact on the replay value of the game.
It is fine to have an AI display different personae as long as it is strong enough. It is not the case right now.

Fortunately the "Match Now" Multiplayer feature provides challenges from players all around the world. There is also a Face to Face feature. All the love and critiques the Agents have gotten in comments, reviews, and this thread really underscore the importance of continued development in both variety and difficulty of Agents. As an advanced player, if you find yourself resorting to a dominant strategy it could be a good opportunity to switch things up. Try to win a new way, or, if you've found a specific exploit, report it and don't use it anymore ^.^

Now, you can leverage that effort by plugging this work into a Monte Carlo Tree Searching system. Let the MCTS search the tree by evaluating nodes by running simulations using your current AI. You don't even need an evaluation function!

To improve strength and response time, build an offline opening book, using the MCTS.

Finally, move the computation to the cloud...

We've barely scratched the surface of what we can do on the iPad. I am also really excited by the opportunities of cloud computing, but it is a requirement that single player be entirely playable offline.

Additionally my approach favors teaching the AI to play the game, rather than relying on brute force computation. It also lets me build up systems that can evaluate the style of a human, and intelligently build models and data to self-modify it's play accordingly. Having played a number of games featuring an offline book, I find the play to feel cold and mechanical. It is my hope that the agents created provide a different type of experience.

Agents have plenty of room for improvement (oh yes, they will be improved), but they will be beaten and that's okay. They play interesting games

<TROLL> Though....If you want to lose so badly, friend me ( Miguel_TSS ) ^.^ I accept Game Center Invites</TROLL>

Have a great Holiday!

[Croix de guerre]

Additionally my approach favors teaching the AI to play the game, rather than relying on brute force computation.

I understand and I respect your approach.

My perception of the Bulge AI has evolved lately. It seems that most issues reported are about the AI making mistakes, especially related to supply chains. I feel now that the AI doesn't need to be smart to be fun. It just needs to avoid obvious mistakes that no human player would make.

I suppose your current approach could be complemented by a system that would veto obvious mistakes, maybe a tree searching system based on a simplified mini-model of the game that would consider only area control and supply.

[IckleFirstie]

If you haven’t seen Miguel’s article on what is behind the 8 AIs, check it out!

Thanks for posting this. Having done a bit of "real world" AI, definitely appreciate the complexity of the problem and your solution to it. I'm a big fan of utility as well.