WEGO WW2 Strategic AI - What Does it Look Like?

[bcgames]

In what ways do you expect a Strategic AI to perform so it will kick your ass 100% of the time? How does it decide on a course of action? What is the criteria? By what measures does it determine how well or poorly it is performing? By what procedure does it determine to continue a course of action or change to another? What does the implementation look like in the editor? What other questions need to be answered to properly construct a strategic AI?

Take a piece. Provide your ideas on potential solutions. No wrong answers here; just ideas.

Go.

[Rosseau]

Strategic AIs already kick my ass 100% of the time.

Sorry, I could not pass that one up. And I'm sure there will be more constructive posts. Your games are already a high challenge for me.

Best wishes!

[bcgames]

Thanks for the feedback. Our goal is to give everybody a good, fun game regardless of experience level. I think the AI needs to serve a broad spectrum of game skill levels. I don't believe we address the needs of the intensely-studied, highly-experienced WEGO grognards that have come to the game. They need to experience a fight that is challenging, satisfying, and appreciated. Adjustable rheostats may be the way ahead. Dunno.

[MrLongleg]

I think creating a good enough AI is the holy grail of computer games, but also very difficult. This game probably is meant to be played PvP. A good AI for a game like this would first have awareness about space and lines of communication. It needs to know which assets need to be protected (e.g. HQ's) and emulate a real chain of command. E.g. each division gets an area of operation, needs to determine the threat level for its area and of course form a reserve.

[bcgames]

This game probably is meant to be played PvP.

I agree 100%. Best memories of playing Stalingrad and Desert War are from PvP. There are some cunning WEGO Players out there! Great teachers of the system. Hats off!

But...~95%+ of the buyers are solo players. I must to give them the best gameplay I can.

A good AI for a game like this would first have awareness about space and lines of communication. It needs to know which assets need to be protected (e.g. HQ's) and emulate a real chain of command. E.g. each division gets an area of operation, needs to determine the threat level for its area and of course form a reserve.

Food for thought. Briefly stated. Nice. Thinking...

[MobPsyduck]

That is certainly an interesting question.

Things I believe the AI could handle better:

When there is a breakthough the AI seems too oblivious about looming threats of encirclement or maybe just not aggressive enough in preventing it from happening. Or if the encirclement is already complete not enough is done to reopen supply lines and/or counter encircle the spearhead.

As a defender I noticed that sometimes the AI brings a handful of tank units together where it wants to break through but then spreads out their attacks too much to make meaningful gains until the tanks are worn down and toothless.

Also the AI seems to like artillery bombardments against my HQs a lot. I have no way of knowing for sure, but I suspect sometimes the artillery would be better used in shifting odds in battles.

Things I believe the AI handles well:

When there are holes in my defenses the AI seems willing and capable of forming smaller encirclements (although a human player would probably act more aggressively).

In my current game the AI is very good at intercepting all of my supply shipments which forces me to think twice about where I should spend what I have.

It was mentioned that the AI should know to protect its HQs and most of the time it does as far as I can tell.


I hope i understood the question correctly.

[benpark]

Considering this game-system in particular -

The ideal is an AI that can approach a given scenario with the allotted forces in a reactive way that resembles a rational use of those forces. This is also taking into consideration the enemy forces/positions on the simulated terrain. Put simply the AI should not attack when it shouldn't, and should not sit tight when it could be reasonably attacking. It should do this all in a reasonably coherent manner, considering local and operational considerations.

The AI also needs to take time into consideration. It needs to know how long it has to accomplish these objectives.

Not so easy-peasy, but this is the ultimate question! If the above is the goal, the AI system of action/reaction should be designed around the concepts above. There should be a balance of getting the AI to behave like a potential game-winner, and another that is getting it do so in terms of adherence to historical factors and how military forces operate in the field.

The AI must be made aware of how to use the types of forces it commands in a basic way. Don't lead an attack, ever with an AT unit if there is plenty of armor and infantry at its disposal. The AI should have this divided into units that have greater attack than defensive properties- this can be "baked-in" to their basic unit stats.

One approach to understanding the map (somewhat akin to the current one, I think but with tweaks) may be to divide a given operational map into different "Zones" or Trigger areas that could be assigned a certain ratio of AI attention in terms of where to attack and where to defend.These Zones would have AI forces assigned- this could be by units, or formations. These units would have a set of behaviors that could be linked to triggers, which then are evaluated for an appropriate action.This evaluation has three passes:

1. Local situation AI evaluation for that Zone (or Trigger area).
2. Operational AI evaluation. This is a weighting of all Zones (friendly and enemy), forces, and force distribution through the Zones.
3. Collation of the two previous steps with decision of course of action for units and sub-units- examples might be like Attack/Limited Attack/Defend/Hold/Delay/Displace for another location. A unit could also have things like aggression and attack/avoid routines checked at this time.

A pre-programmed use of sensible formations for different movements (taking into account formation size) may also help avoid units becoming disjointed.

Example-

German forces are defending central Aachen, US attacking. The map is divided into 10 Zones, with the city center being the most weighted ("Z1"). The two flanking regions are also of elevated importance (Z2, Z3). There are 12 turns.

German unit "A" (let's say a Btln. of sub-units) is in the central location, "B", and "C" in the flanking positions each. Unit "D" is forward (Z4) at the start with orders to Delay the US forces and grind them until it is no longer logical to do so given factors like force preservation and US player taking unforeseen actions.

The game begins, and US forces outflank German unit D. Three turns have elapsed, which Triggers a switch from an AI Delay order to a Displace order. The designer of the scenario could set a weighted number of locations that this formation could then attempt to retreat to. If they become harried along the way, there could be additional settings for aggression and routing. A toggle could also switch to another formation course of action based upon casualties/suppression to something like Defend/Avoid.

US forces now approach the Z2 and Z3 areas of the map (unseen by player, of course). Unit B is attacked, but not C. Once an evaluation is made of the ratio of US to German forces in all Zones for that turn, the AI can release some forces to counter-punch (ONLY attacking ones, and this could be a percentage ratio that can leave the Zone from their Defensive positions if on the Attack).

[bcgames]

Good thoughts. Thank-you for your continued interest in advancing the WEGO game system. Your detailed insights into progressing the AI are much appreciated. Your discussions reminds me of the formation construct found in Take Command: 2nd Manassas and the games that followed. I think Norb Timpko nailed it with his Civil War/Napoleonic game engine. The formation .csv defined where units were located within the line or column of the brigade, division, corps and/or army. It kept the right units of an organization in the front line and others elsewhere.

Couple that with an "intelligent terrain" analysis--and--an organization's formations now having something to adjust to--i.e., avenues of approach, cross-mobility corridors, key terrain, decisive terrain, logical defense lines, and logical axes of attack. All these things would be hard implementations of the military decision making process (MDMP).

At the strategic level, I think the correlation of forces and the status of supply is key to determining the initial posture of forces (Offense/Move-to-Contact/Hasty Attack/Defense/Delay/Retreat/Withdrawal, etc.), and then making dynamic changes of posture based on the changing situation over the length of a scenario.

Thinking...

[gagnradr]

Disclaimer: I didn't study thoroughly how your AI works (yet).

I would expect the AI to work nested routines according to levels of decision making. A Unit is given space, time, forces, and goals, and makes decisions accordingly, breaks space, time, goals down to its subunits and they repeat the process. So an Army breaks 35 rounds, the whole map and 7 goals down into Corps-objectives and -sectors, the Corps breaks that down for its brigades and so on. Once an obstacle is met, the subordinate adapts and reports. For a player, the control of all his small tokens might be fun (though not realistic), but for the AI I think nested routines that correctly represent military decision making are best.

The decision-making would be weighted by the victory points: a destroyed unit is victory point X with a long-term pay-off of f(X) and a campaign pay-off of f'(X), while a victory map area has a value of Y with a pay-off of f(Y). By comparing the costs of own destruction against possible gains, a AI could have a "daring" or "anxious" weighting. And by weighting given goals against possible goals, it could be more or less "mission-type tactics".

Currently, the AI fails to exploit successes and break throughs, abandon continous attacks against strongpoints in fortified lines, evacuate cauldrons, and coordinate sectors (and movements) between different HQs. While it's still great fun, it leads to exploitable behavior (pic attached). Currently, the sudden appearance of enemy reserves and reinforcements behind his lines is greatest feature and great fun.