Concerning the AI

[exsonic01]

Several methods I use to circumvent 'easy AI' issue is:

1) Do manual set up for AI forces, and put more point size to AI, but not too much.
For example, for meeting engagement, I try usually 10k for me, and give 15k~20k, even 25k~30k sometimes for AI. But try not too much number of units for AI, because that would bring easy bottleneck, also there's a chance to game become like tower defense game. Use manual setup, and try to give AI good amount of artillery, proper mix of high/low tier tanks, and mortars/ATGM vehicles, gunships and etc. Try to balance AI army in a way to push in different way. Let them spam smoke shells and DPICMs. This will give you tough challenge. It would still unsatisfying in tactical sense, yet still a bit better experience than see easy peasy piecemeal assault attempts.

This is also somewhat more histrionically accurate. According to Operational Art and Tactics written by D. Glantz, (http://www.dtic.mil/dtic/tr/fulltext/u2/a216492.pdf) numerical advantage, as well as concentrated fire support and proper cover, is one of the key ingredients for PACT forces including Soviet army. Page 8, you can see typical assault operational formation of Soviet army against unprepared NATO defense. When compared to the battalion size NATO defense, the size of OPFOR operational maneuver group is amazing, consist of multiple MRR and TR, TC, TBs on 8~20km wide front. This is almost like ~1:5 numerical difference or even more.

So I always try to give some numerical advantage to PACT when I play NATO. On the other hand, when I play PACT, I set up my forces to have the same amount of point or only slightly more point than AI NATO to have more fun and challenge.


2) Try to choose map with more open field, especially AI-setup zone should have enough open field.
From my experience, if there are too many woods or city blocks or any "difficult" terrain in AI set up zone, AI feels hard to set up attack route, and AI's push become piecemeal attack. I looked via developer mode, and what's happening is that each AI units take different time to pass the difficult terrain, and then the gap among AI units are more increased. Then AI units push to friendly line, and this exactly looks like piecemeal attack. Their push is in order of "who cleared the difficult terrain first?", they push one by one, and wrecked by my tanks.


3) Try to give AI platoons, rather than company based units. This also coincides with 22sec mentioned above.

The AI does struggle sometimes with path finding. You can see this when you watch in Developer Mode. It’s why I don’t purchase company sized formations when creating scenarios. I think is what leads to occasional piecemeal attacks. I would like the ability as a scenario designer to set the AI’s preferred formations and movement path.

Company based formations are huge, and it seems that AI doesn't wisely adjust the formation type and gap among units in the formation. In this case, movement of AI unit become being slowed by so many reasons, just to keep formatino intact. So, give AI more freedom by choosing platoon based forces only. This will help AI to maneuver and push better.





In real army, we use the concept of Attack Position and Line of Departure, and Axis of Advance. See FM 3-90 (https://fas.org/irp/doddir/army/adp3_90.pdf) page 3-4. I know it would be tough, but if it is possible, how about introduce similar concepts to AI?

Check the map -> look for the Attack Position or Line of Departure, by checking ETA to nearest VP and amount of cover on the possible attacking route. -> set up Attack Position / Line of Departure. -> move units to Attack Position or Line of Departure. -> Wait until friendly AI moves near AP or LoD -> Attempt assault.

Position of AP or LoD can be different among units, but amount of time takes to reach VP should be the same or very close among different units. In reality, typically company~battalion units share the same AP or LoD, depending on size of operation. So, for this game, I think it would be OK for AI forces to share 2~4 AP and LoD on the map depending on the game size.

Setting up of AP/LoD for AI would be really hard, but the position of AP/LoD should not have direct LOS from enemy point of view, but arrival time on enemy position should be the same or very comparable. I know it would be really tough, but this feature would make the assault more realistic, nicely timed and orchestrated effort from different direction.

Key here is, ability of AI to set up one or two phase lines and LoD, from various terrains and maps. This will make realistic AI, and solve the current "peacemeal" issue.... at least in some degree. Well, I know it won't be easy, but maybe worth a try?




Also, this is one of the reason why I think it would be great if we have PBEM in the future for this game. Human opponent can bring more headache with creative tactics to any players.



Plus, this will be very long term approach, but how about allowing 'save replay' and let players share the replay? Then you could collect the replay, make a database. Movements, positioning, unit types, geometry and etc... Then you could use those big data, in a way to optimize your AI, using some techniques like ML or genetic algorithm or etc... This would be long and hard way but it would be definitely a good way to improve AI. And this way is kinda already widely accepted in game industry.

Look at Starcraft AI tournament. https://sscaitournament.com/ and you can find some from youtube. It is amazing to see how 'enhanced' AI plays and win against each other. This tournament started from 2011, early AI engines were really bad. But these days it is surprising to watch how AI plans and reacts. I guess this approach will be accepted to any other RTS fields, as well as wargames someday. Then this could realize what would really realistic in wargame, such as mission oriented control.

[Werezak]

In my experience ML techniques are not really useful for game AI. Just keeping this as simple as possible: Applying ML or genetic algorithms doesn't imbue an AI with intelligence, they are just ways of obtaining solutions to a decision problem without prior knowledge... but you need to define the problem first.

IMO the best approach for a strategy game AI is to identify the key concepts used to make decisions and break them down in a way that a computer can understand. The Attack Position, Line of Departure, and Axis of Advance concepts sound like they could be really promising for a game like AB in that respect, but that's not for me to decide.

I have only some experience with game AI personally but this is also what I've seen work for others.


Just as an example I worked on an AI for a sci-fi space 4x strategy game that had a tactical and strategic level. The development that improved the tactical AI the most was just identifying that the tactical game was essentially resource allocation, where the resources were the weapons on your ships (each weapon could fire once per turn) and the allocation was the choice of targets for each weapon each turn. There were a lot of nuances and factors that affected what the best target for a weapon was but the underlying algorithm was a target priority table where the highest priority target in the table was selected. Pretty simple, but it dramatically improved the performance of the AI - the real work was identifying that the underlying concept was key for decision making in the tactical game (and also that other aspects of the tactical game, like ship placement, weren't as important and could be handle with "good-enough" dumb behaviors).

[Rosseau]

ORIGINAL: JamesLxx

First of all STOP. Then either: replan an alternative safer route; or call in artillery/airstrikes; or once knowing what and where the enemy is bring up forces capable of knocking them out.

Show me an AI that does this, and I will buy the game. Maybe Command Ops 2 is close, but certainly not Steel Panthers. Not sure about FP:RS

Play some Combat Mission (latest engine). I see in generated battles, AI is terrible still. It often leads with its BN HQ unit in soft vehicle! That's why the game has a detailed scenario editor to plan AI behavior.

Most games like this, you have to do some creative work like exsonic01 posts above. At least the best games give you some tools to work with.

I am no fanboy, but 30 years and hundreds of wargames, still looking for the holy grail [;)]

[exsonic01]

ORIGINAL: Werezak
In my experience ML techniques are not really useful for game AI. Just keeping this as simple as possible: Applying ML or genetic algorithms doesn't imbue an AI with intelligence, they are just ways of obtaining solutions to a decision problem without prior knowledge... but you need to define the problem first.

I know, a part of my career is related about optimization coding, but not in gaming. But there is a attempt to use ML to improve game AI, like some of the Starcraft AI groups. There are tons of tons of replays from SC1 + SC1 Broodwar during last two decades on SC community and Blizzard community. What some groups trying is that they gather all those replay files, analyze all unit's position, type, movement, map & geometry, build and unit order and etc... from each replays, and make a huge huge data out of them. Then they analyze them, and try to train their AI algorithm using ML scripts, to win the game and to mimic human player's build + strategy + micro + etc... That is why I mentioned about replay above.

However, it is not a easy job and it would need huge computational resources, I don't think this studio or any other small-studios for wargame would be capable of doing that, and I never mean to (Sorry Veitikka I never mean to burden you, I understand your situation and I wish the best for you). Those all comments regarding ML is just purely my wishlist & imagination. But someday in the future, we will see more attempts to develop and train AI using advanced algorithms from usual RTS games and wargames.

IMO the best approach for a strategy game AI is to identify the key concepts used to make decisions and break them down in a way that a computer can understand. The Attack Position, Line of Departure, and Axis of Advance concepts sound like they could be really promising for a game like AB in that respect, but that's not for me to decide.

I have only some experience with game AI personally but this is also what I've seen work for others.

Yes I totally agree with you and this game's AI should also follow. Frankly, I think this game's AI is OK enough, but needs a bit more improvement. (We need to remind that no AI can be 100% creative like humans, at least AI from wargames like this one. So that is why I mentioned about PBEM)

Regarding concepts like phase lines AP, LoD, and AoA, I still think about it but still it would not easy. (Or maybe this game's AI already using similar concept) Those concepts are very geometry dependent. It would be easy to introduce, if we script the single mission for known & decided map. However, this game's unique freedom to choose the battlefield the player want is kinda issue. I'm not sure how to make AI to set their own one or two phase lines, and LoD and AoA in any geometries in this game. It is tricky. For very simple plane map, it would be easy, decision of phase lines and LoD would be just distance from VP or distance from deploy zone. But when there are mountains, cities, and rivers, it wouldn't be easy.

I'm also thinking about what would be the good easy way to make computer understand about those. Key is, ETA to reach to VPs for all assaulting AI units should be the same or similar, once AI decide to start the assault on player VPs or neutral VPs....

Just as an example I worked on an AI for a sci-fi space 4x strategy game that had a tactical and strategic level. The development that improved the tactical AI the most was just identifying that the tactical game was essentially resource allocation, where the resources were the weapons on your ships (each weapon could fire once per turn) and the allocation was the choice of targets for each weapon each turn. There were a lot of nuances and factors that affected what the best target for a weapon was but the underlying algorithm was a target priority table where the highest priority target in the table was selected. Pretty simple, but it dramatically improved the performance of the AI - the real work was identifying that the underlying concept was key for decision making in the tactical game (and also that other aspects of the tactical game, like ship placement, weren't as important and could be handle with "good-enough" dumb behaviors).

That is brilliant work. I also believe that devs will enhance their AI further in the future, after they have enough time and resources.

[Skybird]

On that video example.

What is shown there gives me the impression that the platoons of Red moved in street columns or echeleon, narrow spacing.

Imagine this were a real scenario. Red moves two companies and wants to quickly get to the objective. He does not use streets or highways since there are none, left and right there is hilly terrain or dunes. What tank commander would send his formations in street columns under these circumstances? Street columns need streets to benefit from this formation. - IMO, the proper formation would have been wedge, to provide scanning and fire power to both flanks left and right. Maybe the AI indeed was using echelon right, which might be appropriate due to that small hill nearby, the video does not allow to fully understand whether it was echelon or street column.

Now the first platoon gets wiped out, and from very close by. A real company commander would have become aware of it. Either the tank cannons, or the explosions, the smoke, the ammunition cooking up, the sudden loss of radio communication - whatever. what would he do? One thing is certain: he would be very stuid to just continue in this situation, with losses suffered at closest range, and a slight hill on his right flank, and open terrain ahead.

He would stop for a monent, most likely, AND REGROUP. Form a firing line, or a wedge. Depending on the terrain, sending a scout sneaking on that hill to risk a look over the peak, or sneaking around the hill. Tanks usually avoid going over heights, they usually try to avoid easy detection against the sky by staying low in the valleys betweens hills. Usually.

What would I do? Forming a firing line, and then have scouts flanking from both sides simultaneously. A surprise attack would be out of question, since the enemy knows already I am there, and where. Maybe instead or additionally sending dismounted infantry up to the hill peak.

What does the AI do? It lost one platoon in street column formation. The next platoon rolls in, same formation apparently as if nothign happened, and gets wiped out. And again. And again. Company gone, during most of the cold war, a Sowjet tank company had three platoons á three tanks, and one HQ tank = ten tanks per company.

Realistically, what should an AI be capable to do?

1. Becoming aware that something is wrong. Maybe the game simulates delay in command and control and radio communication. Sowjet capabilties in individually commanding small units "live" were limited at that time, thats why they used huge formations like in Napoleonic wars, and wave doctrine. But they were heavily capable in ECM and supressing NATO radio communication and radar, so maybe NATO would not have been as mobile as was assumed it woudl be... Lets assume the game reflects radio limitations of Red. So his formation will carry on for some time even when getting into trouble.

Wave doctrine implies attack in strength, and combat formation, not street column.

2. Becoming aware of the hill on the right nearby and putting some priority of any kind on running recce on it.

The root cause of the mess is that although there are no streets, Red chooses to let his units advance in a piecemeal fashion, one by one, without mutually supporting each other, in a formation as if there were an infrastructure of any kind. An open plain is ahead. Why not wedge or firing line? Worse, once the mess starts, the AI does not react, it does not regroup, it does not alter its movement vector, it does not react to attack vectors, it does not change formation - in other words: it does nothign to adapt its inadequate stance to the fact that it is under fire. It just assumes that nothing has happened.

After the first fire received, there should have been any kind of reaction, and may it be due to sole self preservation of the individual tank crews following up. The seocnd grouo must have seen whatb ahppened to their colleague at front, the firballs, maybe turrets flying into the air whatever. Instead they follow their dead comrades right in their tracks.

If it were logistica supply units, maybe it would be realistic that they move in column, I don'T know. But combat units, in that terrain, needing to expect battle?


I admit I find this exmaple a bit discouraging, because it does not look different to the aI problem I already saw years ago in the free build. Back then I assumed it would get ironed out, even more so since I think it is pretty obvious a problem. But apparently it hasn't.

Veitikka, I follow your title since many years, always had it on my radar. I communicated its existence in two or even three differen forums over the years. I mean I do not want to anoy you or talk your effort down. It seems there is rich detail and research that was put into the game. But it all stands and falls with the quality of the AI, even more so, for some people at least, when it has no MP (no MP player myself). There are two reviews now that clearly signal that their authors mean it well with AB, but both found that the AI is suffering from the symptom that this thread started about: that the aI anythign but rarely attacks in this alienating piecemeal fashion. So again my tip: before considering spending time on DLCs and additonal features, repair what is of real importance, and that is this AI issue. Its showstopping enough that it makes people like me pausing any buying decision until this is being resolved. For comparable reasons I stopped playing another game in the past which in principle I liked: Steel Panthers MBT. It was much worse there, however. You see, player invests time and meticulously may lay down a plan and executing commands, and then gets an easy victory gifted by the aI lining up to get massacred like Lemmings. This somewhat ruins it. It is not fun, it is not satisfying. Player feel betrayed, he wants to win due to his good plan and exceution, not because the aI "lemmingizes" itself.

The issue is there since the last free build you released years ago. I wonder that it has not been tackled!?

When this gets solved and the game is at Steam, I will become a happy customer. No matter whether it is a bit more expensive at Steam or not.

Nichts für ungut.

[Veitikka]

ORIGINAL: Skybird

Veitikka, I follow your title since many years, always had it on my radar. I communicated its existence in two or even three differen forums over the years. I mean I do not want to anoy you or talk your effort down. It seems there is rich detail and research that was put into the game. But it all stands and falls with the quality of the AI, even more so, for some people at least, when it has no MP (no MP player myself). There are two reviews now that clearly signal that their authors mean it well with AB, but both found that the AI is suffering from the symptom that this thread started about: that the aI anythign but rarely attacks in this alienating piecemeal fashion. So again my tip: before considering spending time on DLCs and additonal features, repair what is of real importance, and that is this AI issue. Its showstopping enough that it makes people like me pausing any buying decision until this is being resolved. For comparable reasons I stopped playing another game in the past which in principle I liked: Steel Panthers MBT. It was much worse there, however. You see, player invests time and meticulously may lay down a plan and executing commands, and then gets an easy victory gifted by the aI lining up to get massacred like Lemmings. This somewhat ruins it. It is not fun, it is not satisfying. Player feel betrayed, he wants to win due to his good plan and exceution, not because the aI "lemmingizes" itself.

The issue is there since the last free build you released years ago. I wonder that it has not been tackled!?

It's totally unrealistic to expect that I can 'repair' the AI opponent to be like a real human. No game like this has anything like it. There's always a reason why our AI does what it does. The better the scenario designer knows the AI behavior the better scenarios he can make. We have a huge variation of terrain in the game, for example the open desert we had in the video, dense forest, rivers, lakes, cities etc. In many cases the AI can advance faster if it uses the column or march formations; for example if it's moving close to a water obstacle in a 500-meter wide line/wedge formation it can become a total mess, when there are just a few narrow crossing points available. These impassable obstacle shapes on the map can be anything: horizontal, vertical, spaghetti. Even an open map can have impassable spaghetti rivers going around there. We have not restricted what the map designer can do, so we must be prepared for everything. It's highly recommended that the map designer knows the engine limitations, and follows guidelines, but this will not always happen. Same goes to the scenario design. In the video, the scenario designer had a huge map, and just a single objective there. This is not necessary good scenario design, but we allow the players to experiment and see what happens. The AI can be made to look very bad, that's for sure. In this case there should have been more objectives to guide the AI better.

Over the years, I don't think there have been many suggestions concerning how the attacking AI should behave, other than these abstractions that the players usually say. Perhaps the reason for this is that the issue is very complex when you really start thinking about it. In this thread there has been at least one good suggestion so far: use the battle formation when under fire. I think the AI does that closer to the objectives. In the video scene there was at least 5 kilometers left to the lone objective.

[Skybird]

I must respectfully disagree in that I think the issue illustrated could be massively improved. Other games show that it can be done.

You probably know Steel Beasts, and that a whole mission could be played from the map screen alone pretty much like in your game. The scenario designer there can set conditions to wapyoints, and he can define in what behavioral mode the unit should travel the path to the next waypoint, and when it should embark, if at all, and how it should move if it does. Also, formation and spacing can be set separately.

Now, ignoring the mission design and focussing on the micro-managing of units' behavior/reaction by the AI in a given situation, a charging unit will get orders by the aI to stay on its ordered path, at maximum speed, racing to the objective, not seeking cover when under fire, seeking no hull down Battle position, only returnign fire. This unit will race on as long as it lives, no matter what.

Normally, the attacking unit will be less kamikaze-like, and have orders allowing it to react to a threat it is aware of. When it gets under fire, immediately it swings turret, often the whole hull, towards the threat, and then will try to find a battle position (hull down) in the vicinity, up to I think 300 or 400 meters around, or pop smoke and back up, trying to hide, or all of that.

So, with the scenario design approach and editor in SB, it can be done. The reaction results of the aI can be very impressive. I admit however that SB has one of the most economic, userfriendly and efficient editors I have ever seen in any such game. Its a brilliant design, even more so since the difference between the mission editor and the map interface used by the player during the mission to command his formations, are not that much different at all.

I do not know what your editor allows scenario designers to do, and we do not even talk about the quick mission editor for the player here. The problem is that the AI in the problematic situations I point at and that can be seen in the above video, does not react at all, no matter what, also, that it has the most unsuitable formation of all, apparently. When the whole force came under fire, it should have quick-jumped to a different formation, and seek battle posiiton as long as not having been given explicit orders by mission designer to not react at all and just race on.

I also recall, although it is a long time ago, Conquest of the Aegean. When the aI in that game , WWII-based and thus using different tactics, came under fire, it reacted. No matter how or what, but it reacted, it did something: changed position, pace or formation apparently, trying to alter its state to aquire a better position in reaction to the thread, returnign fire, calling reinfocrments - it did something. The AI continued to charge into killing zones however - when being ordered to do so by the scenario, like it indeed often happoened in WWII.

Operation Flashpoint, now that game I fear, maybe no other cosim has served me my defeat on a silver plate so often like this one due to the overwhelming firepower of Sowjet helicopters and artillery. Again, when under fire, the AI reacted. It changed pace or direction, it rushed for cover, or it hid behind a building and sat still, later staging a pop up and ambush kind of thing, for exmaple. It reacted to the thread. It still went for the objective, may even charge for a close-up fight. But not in a way to be seen in this video example. Also, the msisions are desiogned usually to allow the AI to attack in force, not in a piecemeal fashion. Result: I often ha da good plan, managed to keep suffieint units of mine alive, acted mobile all over thre place or hunkered down, picking Red units in masses: and still ended getting chewed up.

What the two reviews mention, and what the video illustrates, and what I describe is that your AI in certain circumstances, and apparently not rarely, simply does not react at all, just carries on its business as if nothing happened. Also, starting to move in probably inadequate formations may be an issue here.

The only coding of stuff I ever did, is thirty years ago, and did not go beyond getting a graphical interface for chess and having the computer moving pieces according to the rules. Further I did not get, and real life got in the way, and interest waned, and such is life and blabla... So I cannot tell you specifically how to code such stuff. I only point out that it obviosuyl can be done, somehow.

You have a great game there, with plenty of depth and detail,obviously, much time invested, years of your life. But I admit this deficit with the AI really is holding it back, if it is like it was years ago, the last time I made experiences with it. Back then I won battles by finding good psoitions for my units, and just waiting, if the other side not coming, trying to bait them with a fake attack of a small unit of mine that easily withdrew again. Most battles ended in this way, that I just waited for the Lemmings to pop up even if they were in defence positions. The longterm motivation sufferd very quickly for me, and I cannot imagine I am alone in that. After maybe a dozeh b attles done or so, I saw no point in doing another one. An AI making the gameplay suffering from always the same Lemming-event, is a really serious issue.

The AI needs to learn how to react more properly instead of just marching in street columns while ignoring all events around it.

I could imagine that the combination of if...-then... triggers with behavioral conditions that SB allows to set for both single waypoints and individual legs linking them, could be somethign giving good ideas. I know SB since its beginnings, although in recent years I have not played it that much anymore (Im 51, interests are changing a bit...). A well designed mission can work wonders with this editor's tools which are surprisingly easy to handle. See if you can find some ideas in that system! That is the only practical tip I can offer. Also, SB demonstrates that the micromanaging of the behavior of a unit coming under fire can lead to better reactions than in this video. Occasionally, of course the SB AI will mess it up. Every AI does so at times, nobody demands that it must always be fail safe in its performance. But the failings should be reduced to an utmost possible minimum.

Give it some thought, if you have time and interest left! But just doing DLCs with additional units and missions, will not really improve the issue, for it then will still exist.

[Veitikka]

It seems you're missing the important point here, if you compare AB to SB: the AB AI is fully 'dynamic', whereas SB is fully scripted. That is a very different world. The last time I tried SB, the scenario designer had to carefully script all formation movement. There could be a medley of these two approaches, but for example Close Combat and the old Combat Mission games didn't have any scripting features.

How the individual units should behave under fire, well that's more or less a separate discussion.

[kevinkins]

[/quote]

No game like this has anything like it.

AI is a difficult challenge for developers of detailed war games. Today the only way to have human like AI that I am aware of is to involve the designer in scripting/couching the Tac AI. This impressive feat was pioneered by Battlefront years ago with the introduction of their CMx2 engine and the AI Editor.

From the engine manual:

"The scripting aspect of CMx2 is also critical for coaching the Computer Player on how to win. For example, if the defending Computer Player needs to defend a set of buildings inside a city, the designer can specify which key spots should be manned in order to ensure the success of a more abstract victory condition. No matter how good a dynamic AI may be, practically speaking it will never equal the insight the designer has without some very specific help. Therefore, think of the scripting as the designer helping the AI understand what it needs to do, where," (when) "and how."

Interested players can Google "combat mission engine 3 manual" and go down to page 94 to see the details. Scenarios can be historical or hypothetical or generated via the Quick Battle process. Quick Battles are the battle generator and a landmark idea. Quick Battles are fought on QB maps which use the same "coaching the Computer Player" concept. Hundreds are released per commercial product and many more are community created.

Here is a tutorial that brings the manual to life:
https://www.youtube.com/watch?v=f02X2cygzFY

The effort to create "human like" AI is tremendous given the resources available to today's war game developers. Having human-like AI also means a lot more time and effort into the scenario design process. A good well tested CM scenario can take a few months to release. We need to be careful of what we ask for. Some day a lot of interesting games may never see the light of day. CMx2 should be consider an exception, not the rule.

Kevin

[Skybird]

ORIGINAL: Veitikka

It seems you're missing the important point here, if you compare AB to SB: the AB AI is fully 'dynamic', whereas SB is fully scripted. That is a very different world. The last time I tried SB, the scenario designer had to carefully script all formation movement. There could be a medley of these two approaches, but for example Close Combat and the old Combat Mission games didn't have any scripting features.

How the individual units should behave under fire, well that's more or less a separate discussion.

You are right on SB, but I am aware of it, thats why I separated the mission design aspects from what I called the micro-management of behaviour by the AI. The latter is always present, in any form, even if you just set up your platoon and an enemy platoon and fire a shot at it in this simple design, without further conditions applied, the AI will react: seek BP, if not possible, face you with the front of turret and hull. It gets aware of your presence. Mission designer can, but must not overrule its autonomous reaction. But a default reaction is always there. At least last time I played it, which is two years ago or so. Or an AI unit with a scenario-design of following a path with normal marching or engaing orders to a waypoint: when it gets under fire, it may stop, seek a battle position up to 300m or more away, and fight - that is not scripted, the choice of which enemy it faces and where it sets its BP and what enbemy it fires first at, is done by the autonomous AI, not the mission designer. So, both scripted events and conditions and autonomous AI work hand in hand. If the path has assault orders, the unit will behave like in the video above, but have its turret pointed at the last realised direciton of threat. Death race tactic.

This - I call it autonomous - reaction by the AI, and scripted mission design and its commanded reactions, are two different things. But they work together, can be combined.

I admit and I agree with Kevinkin, however, that coding a good AI I imagine to be the most difficult task of it all. These days I am busy with serious simracing for the most, and I see AI fails and tiumphes a lot there. And I saw how long it may take a developer to get his AI into a state that does not frustrate players with its flaws and incapabilities, but actually is a worthy thing to race against. Again, utmost perfeciton must be avoided as well, for that ius not human: having a human-like AI means having it making errors: just not too often or too rarely, not to serious while and not to irrelevant, but just about beign "right". A very difficult
balancing.

Its a difficult task, no doubt. And needs a lot of trial and error to find out about, I imagine. Ideally you have a neural network and have it learning by feeding it with hundreds of thousands and millions of matches played between competent human players. But that is almost certainly out of range for AB, or not?

[noooooo]

I think the an easy solution is to add a "Seek cover on contact" SOP. The defend functionality is already there to automatically find cover, so all the SOP has to do is automatically use "Defend" order on contact with the target being the contact enemy unit.

[Werewolf13]

ORIGINAL: Rosseau

I am no fanboy, but 30 years and hundreds of wargames, still looking for the holy grail [;)]

I'm with ya there, Rosseau. Been searching for 39 years and haven't found that grail yet. After adding up the cost of that search I wonder if that's the way the game makers want it to be. A lot more money to be made in the search than in the finding.

[Werezak]

A neural network is not going to produce a good AI for a wargame.

In order to have any success with a NN you would need to answer many of the difficult questions before you can even set up training. In other words you would have to create a good wargame AI first before you would be able to have a good wargame AI that incorporated NN. The same applies for genetic algorithms, reinforcement learning, or any other ML technique.

Please stop bringing up ML with regards to the game AI. The most productive thing to discuss would be the CONCEPTS used by people in the real world to conduct military operations that would apply to the game, as well as the CONCEPTS used by experienced players.

[exsonic01]

ORIGINAL: Werezak

A neural network is not going to produce a good AI for a wargame.

In order to have any success with a NN you would need to answer many of the difficult questions before you can even set up training. In other words you would have to create a good wargame AI first before you would be able to have a good wargame AI that incorporated NN. The same applies for genetic algorithms, reinforcement learning, or any other ML technique.

Please stop bringing up ML with regards to the game AI. The most productive thing to discuss would be the CONCEPTS used by people in the real world to conduct military operations that would apply to the game, as well as the CONCEPTS used by experienced players.

Maybe it was my mistake to bring ML in this post XD I'm sorry about all those hypes. I feel responsibility to ruin discussion XD

But it is possible, I mentioned in above post about SC. However, it would not only need so many conditions, but also requires huge amount of data, and very nice computational resources. So, it would be more proper to say 'impractical' for this game or any other small game studios at current stage. I also agree that we should not discuss with ML or NN in this post any further.

[Skybird]

ORIGINAL: Werezak

A neural network is not going to produce a good AI for a wargame.

In order to have any success with a NN you would need to answer many of the difficult questions before you can even set up training. In other words you would have to create a good wargame AI first before you would be able to have a good wargame AI that incorporated NN. The same applies for genetic algorithms, reinforcement learning, or any other ML technique.

Please stop bringing up ML with regards to the game AI. The most productive thing to discuss would be the CONCEPTS used by people in the real world to conduct military operations that would apply to the game, as well as the CONCEPTS used by experienced players.

Still this is how Google Team Alphas's project created the strongest Go engine out there that plays against the human world elite - successfully. Backgammon was solved this way as well already years ago. I think Google in case of Go had its machone not even knowing the basic rules of Go, but simpy had their code core observing the going of Go macthes - MANY of them. The patterns of rules and later strategies the machine formed by these observations by itself. The rules were not "explained" to it or coded into it. It concluded on them by itself, by watching matches.

If you want an AI to recognise an animal, you show its core algorithm managing the neural network pictures of that animal, in a thousand different situations, in all kinds of habitats, in all light and darkness situations, in full display, in cover, in flight, sitting on the ground, in colour, in black and white. Hundreds of thousands of such pictures. At some point the algorithm has so many data sets for comparison that it will recognise that animal when its optics see it, and will do so under all circumstances.

Thats the basic principle by which you teach a neural network: You let it observe an immense amount of events, objects or whatever it is about.

Assuming you have such a neural network and you have a database with 3 million battles played in Armored Brigade. You have an interface by which the machine can observe these matches, one by one, one after the other. You must not explain it the rules and what options there are. If the number of matches that serve as demonstrators is big enough, it will conclude on the rules and options' meanings and then on successful tactis and strageies by itself due to the repeation of things. Like Google's Go machine understood all by itself how Go is played and what the best strategy is and how to calculate that far ahead in moves that you can compete against world class players - AND BEAT THEM.

As one neurla networkj expert some months ago said in a documentaiton I saw on the matter: you teach a neural network by letting it watch a very high number of observations - in principle that's it.

Simple ,eh? [:D]

[vorbarra2]

That's true, but not useful. You can teach a computer how to recognize a cat with that type of learning, and therefore any other type of image recognition, but thats a binary outcome. Is the image cat? is a yes/no question. In the case of Go, though the board state is complex, there are still only a finite number of states the game qcan can exist in, snd the only imput space for the AI is 'where do I put the counter?'

AI for games with an essentially unbounded number of internal states (ie. all computer game design as we know it) does not conform to this type of learning. The number of instructions the AI has to process is very large, it's not 'the counter goes here,' but many different types of instructions which don't map neatly to the binary outcome classification of machine learning. The best that can be achieved is a generalized algorithim, which probablys outperforms the majority of human players (certainly if my record is anythign to go by) The comments by the dev are trying to convey this point, which is basically 'so how would you express that in a symbolic, generalized fashion that would apply to all possible edge states and conditions and unit choices?'

Things like pathfinding and pathing though are more solvable and I think the way in which units path could certainly be improved. It;s not that obvious to me why they take the path they do, and the interface does a poor job of telegraphing what commands would be needed to ensure a certain type of movement. I can formulate the thoguht 'mech inf, stick to roads, get to the town fast, chuck out your passnegers and take up defesnive positions in or behind buildings,' but the pathing or formation commands never seem to make it easy to achieve this.

I wonder if the comments people are making about piecemeal advances are really an issue of tactical-level decision making, or simply poor pathing leading to formations getting broken up and encountered piecemeal? Does the AI have an internal metric for checking formation integrity?

[noooooo]

ORIGINAL: vorbarra2

That's true, but not useful. You can teach a computer how to recognize a cat with that type of learning, and therefore any other type of image recognition, but thats a binary outcome. Is the image cat? is a yes/no question. In the case of Go, though the board state is complex, there are still only a finite number of states the game qcan can exist in, snd the only imput space for the AI is 'where do I put the counter?'

AI for games with an essentially unbounded number of internal states (ie. all computer game design as we know it) does not conform to this type of learning. The number of instructions the AI has to process is very large, it's not 'the counter goes here,' but many different types of instructions which don't map neatly to the binary outcome classification of machine learning. The best that can be achieved is a generalized algorithim, which probablys outperforms the majority of human players (certainly if my record is anythign to go by) The comments by the dev are trying to convey this point, which is basically 'so how would you express that in a symbolic, generalized fashion that would apply to all possible edge states and conditions and unit choices?'

Things like pathfinding and pathing though are more solvable and I think the way in which units path could certainly be improved. It;s not that obvious to me why they take the path they do, and the interface does a poor job of telegraphing what commands would be needed to ensure a certain type of movement. I can formulate the thoguht 'mech inf, stick to roads, get to the town fast, chuck out your passnegers and take up defesnive positions in or behind buildings,' but the pathing or formation commands never seem to make it easy to achieve this.

I wonder if the comments people are making about piecemeal advances are really an issue of tactical-level decision making, or simply poor pathing leading to formations getting broken up and encountered piecemeal? Does the AI have an internal metric for checking formation integrity?

Took the words right out of my mouth. When you look a go board, say a 19x19 board, you have 361 possible places to put your stones. And less over time as the board becomes filled. Even that creates a enormous amount of possible permutations which is why AlphaZero (chess AI although chess has much less permutations than Go) and AlphaGo is so impressive.

But if you look at something like AB or other computer games, the amount of possible states outnumber Go by such a massive amount that it's not even close to comparable. The fact is, when a tank is pointing 56 degrees northeast is technically a different state than if everything else being the same the tank point 57 degrees. This is talking about ONE tank and ONLY the direction it's facing. It's simply not even remotely on the same magnitude as something like Go, and Go is ALREADY a massive undertaking.

Using ML for something like this is not the equivalent of recognizing a cat or playing Go. It's like asking it to speak English to the level of an average human. There are so many contexts and concepts that there simply isn't enough computing power or resources to do it.

[kevinkins]

As an avid chess player, I have followed the AlphaZero story in amazement. AlphaZero AI beat the champion chess program after teaching itself in four hours. It accomplished a similar thing with Go. I believe another AI has become very proficient a poker. Both Chess and Go are complex I go you go with a single piece game. These milestones are really used to demonstrate of far neural networks have come. I believe the technology behind these demonstrates are really meant to tackle more important problems. That being said, it would be fascinating to see how a technology like AlphaZero tackles a basic ground war game (infantry, light and heavy armor and light and heavy arty) on a mixed terrain map using a basic rules set. Don't' give it 4 hours, give it a day or two. I have no idea how you would set up the training. Maybe a meeting engagement with map exit and the objective. Would the tactics developed mirror human tactics? I do not know. What rules set would be a good to start with? Something from the Tiller genre or more simple like PanzerBlitz.

Kevin

[vorbarra2]

No it's not an issue of time or training sets. The AI instruction sets for something like AB cannot map to the type of ML framework used by the Go or chess learning tools. It's a different type of computation trying to solve a different (limited, ridigly defined) type of problem.

What you are thinking of, is I think, more like 'what is the consensus-best flowchart that can be generated which we could implement into a generic algorithm for the computer?' That is a different type of question to what the ML platforms achieve.

[maxb]

seems to me the easy solution is to have multiplayer. the game seems cool enough and easy enough to pick up that multiplayer would be a no brainer.

it's definitely better than Steel Panthers, Close Combat, and FC:RS in my humble opinion. It's like a mix of those but superior.

nobody would whine about AI if there was MP