A question about AI architecture?

[MarkShot]

Please forgive me for always bringing up other titles in the forum, but I am truly trying to understand THIS game.

A user by the ID of Kull explained to me on the WITP-AE forum that this is how the total AI works:

* There is a generic AI. So like we see. "Forming front line" "Moving into vacated hexes"

* However, there is also a strategic AI for every included campaign or scenario which is a script; known as a variant. That there may be as many as 20 or more. And one is selected at scenario launch.

So, I was curious is WITE-2 also a hybrid of hard coding and scripting (variants)?

If there are variants, given the enormous nature of the 41GC, what does the scripting portion of the AI do when things are not going to plan?

When we were working on the Panther Games series, we had a 100% generic AI without scripting other than the use reinforcement schedule and influencing behavior with VLs (like we would simulate securing a supply line with a trail of low point bread crumbs (along a highway) approach as the AI was not supply aware). So, often an attack might get stalled (based on the losses and rout status), and an AI replan would be propagated throughout the side to reevaluate resources and goals. The FUP|axis of the attack might change. An attack might break into two axis. A decision might be made to bypass and look to cross a river else where. ...

So, what does WITE-2 do?

And I am also thinking of our AI we for the AGE engine of AGEOD. In general, there was no continuity across plans. Every turn (WEGO) was a complete reevaluation of the map situation and a response to it.

So does WITE-2 form plans that persist for longer than a single turn?

Finally, beyond rules regarding OOBs and units is doctrine represented in the hard coded portion of the AI? (As I find most games have a single AI regardless of which side it plays.) (Which was true of PG. We discussed doctrines and it has always been on the list.)

Thanks.

PS: Don't give away any spoilers, but I find the AI to be quite good for a game, and am curious how you did it?

[Joel Billings]

Yes, there is a basic "campaign" AI that is generic. There are lists of objectives for each side and I think Gary has them set up by Axis Army and Soviet Front. There are a few decision points, for example deciding where the main strategic effort will be for the Axis in 1942. There are different AIs for the attacking and defending side (re how it tries to form and maintain lines, how/when it will attack, etc.), and I think there are some differences between the Axis and Soviet AIs. This generic AI applies to all scenarios. Some scenarios, mostly for expansion scenarios we're working on now, have some specific AI written for them. The editor allows a flag to be set indicating the specific scenario, and Gary uses this to hard code some specific AI code for that scenario. Sometimes he starts writing some specific code, and then if he likes out it is functioning, he'll add it to the generic AI. In that way, the expansion scenario work gives us the opportunity to improve the AI over time. This has happened since release as one of the scenarios being worked on involved the Caucasus offensive in 1942. Gary wrote some special code to handles parts of this, and then integrated some of the code into the generic AI. That's the extent of my knowledge of what Gary is doing.

[MarkShot]

Thanks.

I think the part which impresses me the most is back propagating successful strategies to the core AI. Not exactly machine learning, but a human-machine hybrid. EXCELLENT!

[Teo41_ITA]

Outstanding! I was about to open a post just to praise the developers for their outstanding and magnificent work in setting up the AI. I will be more than happy to buy every future expansion of this game, as every cent invested into it it's definitely worth it! I am having a TON of fun against the AI on steroids (= impossible difficulty) and that's what I LOVE about this game!

[MarkShot]

I have a WITE-2 game of Road to Leningrad to which I must return.

Previously, I was testing out WITW an approach of not tethering armor to an advancing rail line, but instead supplying it via air freight by take an airfield every turn. The approach worked well for speed, but of course the allies have lots more air freight in WITW than the Germans in WITE-2.

The strategy was working well in WITE-2 until in a single turn the Red Army formed a loose noose around me. I don't think its fatal other than the clock, but boy was I impressed. Why? This didn't look like an algorithm that simply got lucky. It appeared very purposeful in its actions. It did not fire a single shot and waste strength banging away at me. Instead it very expeditiously closed a noose and isolated my units. If I don't react immediately it will gain the upper hand and be able to do some real damage.

For an AI that gives the impression that front integrity is a driving principle, it switched very rapidly from what what in chess would be called a pawn skeleton to a positional game.

Joel (please convey to Gary) ... so often one buys games which are marketed for their vast scope (insert PDS here as an example), and yet the engine itself cannot actually play the game the designers built. Well, I didn't get the feeling at all that I was wrestling with a digital cripple; no, not at all!

Cheers!