Ideas for helping AI

[donaldseadog]

With the new abilities provided by Tony via the exportDLL functions I've recently pondered how can we level the playing field a bit for poor old AI? Any ideas?

What do people see as AI's short comings?

[CV32]

There are plenty. I'd suggest organizing or categorizing the areas of concern, e.g. sensors, navigation, etc, etc.

[CV32]

One mechanism that would have potentially immediate good results for both the AI and the player, for that matter, would be to enable attacks to be launched from a scripted patrol point.

 

Not just attacks that the scenario designer thinks will be possible (because a player group can be detected there, although these are useful), but scripted attacks launched from that patrol point against a known player target.

 

This would go a long ways toward permitting off axis attacks, i.e. unexpected attacks that do not come straight down the bearing from known AI controlled bases and groups.

[TonyE]

One mechanism that would have potentially immediate good results for both the AI and the player, for that matter, would be to enable attacks to be launched from a scripted patrol point.

The AI topic is so huge I always just freeze up and can't write anything at all so thanks for kicking it off Brad, I'll piggy-back.

 

I feel air strikes is an area where the game has enough structure for the AI to do a much better job. Don has already written code to modify paths. He has the ability to tell the AI to put together multi-axis attacks without a scripted patrol point. Time on Target (ToT) also isn't easy but all of the information is there now that database access is available. With a good bit of work the AI could even set up its own refueling tracks for a package of strikers, not to mention get the SEAD birds on target before the heavy-hitters.

 

Air operations are low-hanging fruit imho, big gain for modest effort. Multi-axis attacks by the AI without coordinated ToT might be a decent place to start.

 

In the big picture the AI needs a strategy. The current potential inputs to that strategy are victory conditions and the AI's order of battle. The VC pillar is weak (VCs do not describe in enough detail the scenario goals). An entirely missing pillar is starting intelligence about the opposition. The AI needs quantifiable information about the expected threat, whether accurate or not, from the scenario author. Building the strategy engine would require coordination between Don and myself since it requires changing & adding many game structures. From the strategy all things would then flow. A quick example would be Operation Island Wind in WestPac. The North Korean AI would need to know its purpose is to hold off the technological superior forces from taking out the missile launchers before the missiles can be launched. It would look at its assets and decide that it has to put targets (planes, as ill equipped as they are) in the air to delay the attackers despite horrendous loss because launching the missiles is the most important thing in the scenario.

[donaldseadog]

I think Don would need some hints

[Palex80]

It could also be helpful if the AI had an idea about the value of each target and the forces allocated at it.

I often see the AI target SAM installations around an airfield and completely ignore the airfield. The same happens some times with ship groups with huge repeated volleys of missiles launched at only the picket ships, ignoring the high-value target.

[CV32]

It could also be helpful if the AI had an idea about the value of each target and the forces allocated at it. I often see the AI target SAM installations around an airfield and completely ignore the airfield. The same happens some times with ship groups with huge repeated volleys of missiles launched at only the picket ships, ignoring the high-value target.

 

Much of this is because the AI is attacking the first player unit it detects.

[donaldseadog]

I'll start looking at what is exposed via the exportDLL functions and see if there are some easy tweaks - better target value evaluation might be a goer.

[Palex80]

 

It could also be helpful if the AI had an idea about the value of each target and the forces allocated at it. I often see the AI target SAM installations around an airfield and completely ignore the airfield. The same happens some times with ship groups with huge repeated volleys of missiles launched at only the picket ships, ignoring the high-value target.

 

Much of this is because the AI is attacking the first player unit it detects.

 

I understand, however perhaps one could install a mechanism to tell the AI when to stop attacking a specific unit.

One of the situations one often experience is that a SEAD-package of 6 aircraft each armed with 4 Harm missiles is launched against an airbase. The airbase consists of the base itself, 1 long rage radar, 1 medium range SAM battery and 2 shirt range batteries. The SEAD package then launches all Harms in groups of 2 missiles over and over again at the first closest unit it detects, for example one of the close range SAM batteries. All other units are ignored, because the SEAD package loiters as soon as the closest part of the group enters the range of the Harm missiles.

 

There are two possible workarounds for this:

 

1. Prevent the AI from repeating attacks, that have not ended yet. The AI tends to launch weapons at the target before the already launched weapons are still airborne. Perhaps one could tweak that, so that the AI waits until a) the weapons have hit the target or have been shot down themselves?

2. Prolong the time an attack takes place at the same target. Perhaps prologing the time between weapons' launches would be an alternative (although a very inflexible one). For example the AI needs to wait 2 full minutes before relaunching at a target already attacked by the same unit? This may however cause significat problems in air combat, so I would restrict that to air-surface & surface-surface attacks.

[CV32]

I understand, however perhaps one could install a mechanism to tell the AI when to stop attacking a specific unit. One of the situations one often experience is that a SEAD-package of 6 aircraft each armed with 4 Harm missiles is launched against an airbase. The airbase consists of the base itself, 1 long rage radar, 1 medium range SAM battery and 2 shirt range batteries. The SEAD package then launches all Harms in groups of 2 missiles over and over again at the first closest unit it detects, for example one of the close range SAM batteries. All other units are ignored, because the SEAD package loiters as soon as the closest part of the group enters the range of the Harm missiles. There are two possible workarounds for this:

 

1. Prevent the AI from repeating attacks, that have not ended yet. The AI tends to launch weapons at the target before the already launched weapons are still airborne. Perhaps one could tweak that, so that the AI waits until a) the weapons have hit the target or have been shot down themselves?

2. Prolong the time an attack takes place at the same target. Perhaps prologing the time between weapons' launches would be an alternative (although a very inflexible one). For example the AI needs to wait 2 full minutes before relaunching at a target already attacked by the same unit? This may however cause significat problems in air combat, so I would restrict that to air-surface & surface-surface attacks.

We've been tweaking the AI for a long time, and its a real chore to find a behaviour that provides good strategy in one respect without compromising good strategy in another.

 

Problem with #1 above: it lends itself to piecemeal attacks, which might be easily defeated by the player.

 

Problem with #2 above: the same result as #1, unless ...

 

I think if the AI could somehow be able to assess the "value" of the target(s) it was attacking (and not just its DP value, though that might be a factor or consideration), AND hold fire until it detected more targets (e.g. an entire surface ship formation, instead of the picket) and determined their value.

 

Something akin to the manner in which the AI holds fire for air to air attacks until the target enters a sort of 'no escape' zone.

 

[broncepulido]

 

AND hold fire until it detected more targets (e.g. an entire surface ship formation, instead of the picket) and determined their value.

Perhaps a very simply mechanism, like force the AI to retain fire by 5-10 minutes? (and in this time-lapse probably had detect more targets?)

 

Another iteration, the retain fire time-lapse randomly variable, or customizable.

 

And if the AI order to open fire is time-randomized you will experiment the thrill to wait for when the AI will open fire against you!

 

Another variation, the AI order of open fire or not could be also a random variable or customizable (but I think my first two ideas are better and easier to implement).

[CV32]

Perhaps a very simply mechanism, like force the AI to retain fire by 5-10 minutes? (and in this time-lapse probably had detect more targets?) Another iteration, the retain fire time-lapse randomly variable, or customizable.

And if the AI order to open fire is time-randomized you will experiment the thrill to wait for when the AI will open fire against you!

Another variation, the AI order of open fire or not could be also a random variable or customizable (but I think my first two ideas are better and easier to implement).

I like the idea of the random variable, though I expect the time to engage would have to be much shorter than 5-10 minutes.

 

Probably taking into account an average of the cruise speed of most bomber/strike aircraft?

 

The AI would also have to continue to close the distance to the target.

[broncepulido]

Oh, my idea of 5-10 minutes time-lapse did come because I was thinking in the detection by part of the AI of one of our warships groups, i.e., as when the recce plane detects a single ship, but some few minutes after detects other enemy ships in the same formation.

 

I was not thinking of air-to-air interception, perhaps we need two algoritms in the AI for that, one faster than other (the faster for incoming aerial targets). Perhaps two different algoritms one for sea targets and other for air targets is too complicated.

[donaldseadog]

What do people think about AI "knowing" victory conditions and using that information to influence its play?
I haven't thought much about it and don't know how it would be done (I think it could) but I like the idea of AI having some sort of ranking of importance regards its targets.

I'm still plodding along looking at some simple games and what AI is doing but not yet learned enough to try out anything.

I'm also wondering about how people approach situations were a new contact is made, what do you do?

[broncepulido]

After a new surface contact I wait a while to see the almost complete enemy force disposition, and only after to arrange any attack to selected target.

The problem is in the same situation the AI fires almost everything at the first contact, instantly, and depleted his magazines against a single enemy ship.