Experimental AI

Other talk that doesn't fit elsewhere.
This is for General Discussion, not General chat.
User avatar
Rman Virgil
Professional
Professional
Posts: 3812
Joined: 25 Sep 2006, 01:06
Location: USA

Re: Experimental AI

Post by Rman Virgil »

------------->

* I hope this is not distracting.

* One of my main reasons for participating is to continuously extend my learning. :)

* I understand Classical Conditioning in biology as a form of associative learning leading to altered behavior. But in bioogy it is most definitely linked to the "survival instinct" and "reinforcement rewards". The Pavlovian demonstration is purely elegant.

* My pondering is along the lines of what exactly is the code equivalent of the "survival instinct", the "reinforcement-rewards" and the mechanism between the 2 leading to some sort of emergent and generative intelligence in the sphere of mixed arms combat.
ilTallman
Greenhorn
Posts: 8
Joined: 04 Jun 2008, 21:28

Re: Experimental AI

Post by ilTallman »

Troman wrote:To be honest at the moment I can't imagine how you plan to apply CC for the higher levels, since usually it is applied to individuums, as opposite to group of individuums, but I have no practical experience with CC so maybe I'm missing something.
A couple examples would do...

1) The bot attacks a Tower with infantry units (excuse me if I do not have all the WarZone 2100 jargon down yet), and finds that his force get shot up. Next time he attacks with Tanks and does well. CC would reinforce this action so that Tanks would be used to attack buildings et al'

2) The bot makes an alliance and finds after awhile that the alliance gets broken and he is forced to loose ground because he was not expecting it. CC would reinforce that alliances should be met with caution.

3) The bot attacks in force, with mixed units, and finds that his advantage was not as great because he attacked with a 2-to-1 force. CC "could" provide a respond in force response so that his next time he attacks with 3x1 or 3.5x1 what ever the bot learns. This has a dynamic effect that bots will attack with more caution and enrich the game play. The same holds true in defensive roles as well.

4) More of a in-Game adaptation, where the bot learns to avoid specific routes because it takes too much time, does not provide tactical advantages etc. So the bot adapts to the map and find good attack routes.

...before you say wow! what a hugh task, remember we are not programming all this to happen. The technology would be set up to handle data complimentary to this environment with the real-hard-core stuff within the provided technology learning these "Rules" or "Responses". We would just have to set up the solution. An of course, its always harder than you think.

Another example would be to learn what sequence works best in the research area. By virtue of playing many styles of these types of games - and indeed somewhat realistic - the sequence or priority of the R&D really makes a difference giving you advantages when you need them. Learning that another advance in weapons does not compare to Intelligence at a given point in the game may pay big dividends.

All this is somewhat speculative because I do not know that much about the game yet. I am working through the code at a snails pace. But the idea is feasible.


Hope that helps!

:)
ilTallman
Greenhorn
Posts: 8
Joined: 04 Jun 2008, 21:28

Re: Experimental AI

Post by ilTallman »

Rman Virgil wrote:I understand Classical Conditioning in biology as a form of associative learning leading to altered behavior. But in bioogy it is most definitely linked to the "survival instinct" and "reinforcement rewards". The Pavlovian demonstration is purely elegant.
CC is only one of the basic learning tenants we have. There are others like Latent learning, Associations etc which do provide functional learning. I lobe Pavlov because although its basic, I think it accomplishes more than we give it credit for.
Rman Virgil wrote:My pondering is along the lines of what exactly is the code equivalent of the "survival instinct", the "reinforcement-rewards" and the mechanism between the 2 leading to some sort of emergent and generative intelligence in the sphere of mixed arms combat.
Great Question. To learn something, there MUST be some trigger to provide the pain/reward. In this example, the system would use context interrpretations to provide insight into its metric of how well its doing. Or simply, using "Won Battle" = "Positive" trigger, "Lost" = "Negative" one. Prolonged Survival during game play provides confidence, alliance length same thing. Loosing a base or building is a big negative etc.

These triggers could then be used to reinforce the data coming into the mechanism to provide the responses. But, I must stress, that the early versions of this would have a lot of manual training. Later models could evolve with instinctive responses - say a database of known responses to situations. Thats gets a little more tricky, but again all doable.

The real hard part is changing the current game technology to something that could interface with such a system. The interface would be like a chalkboard status report (or something like that) providing status of current events. You would even organize it based on areas of the map. Results would be the commands which are layered back to the Strategic<>Tactical layers discussed earlier. That way, all tactical commands provided to local units fit the Grand Strategy - all this happening remotely in my thinking.

Not a distraction at all xD
User avatar
Rman Virgil
Professional
Professional
Posts: 3812
Joined: 25 Sep 2006, 01:06
Location: USA

Re: Experimental AI

Post by Rman Virgil »

--------------------------->

* Hmmm. I c. Very interesting. Thank you for taking the time to explain ilTallman. :)

* I think it might be helpful for me if I look at the "EINSTein" Prog source code and it's hefty documentation: "Artificial War: Multiagent-based Simulation of Combat" by Andrew Ilachinski again.. it's been a couple years since I last went thru it all and I sense a correlation here.

* Thanks again.

- RV :)
User avatar
Rman Virgil
Professional
Professional
Posts: 3812
Joined: 25 Sep 2006, 01:06
Location: USA

Re: Experimental AI

Post by Rman Virgil »

------------------->

* "Brain 4" would be a terrific tool for base and extensible concept mapping here (pre-coding), IMHO.

* After a couple hours spent re-familiarizing myself with "EINSTein" (proggy, source and documentation) I believe it would also be quite useful in this context.

- Cheers, RV :)
User avatar
Molotov
Trained
Trained
Posts: 34
Joined: 07 Jul 2008, 23:22
Location: Texas

Re: Experimental AI

Post by Molotov »

Rman Virgil wrote: My pondering is along the lines of what exactly is the code equivalent of the "survival instinct", the "reinforcement-rewards" and the mechanism between the 2 leading to some sort of emergent and generative intelligence in the sphere of mixed arms combat.
This is, quite possibly, the most elegant sentence I have real all year.
You have five minutes to comply,
0R 3|$3 1T$ FR%1N6 T1M3!!!!!11!!!1!
User avatar
Rman Virgil
Professional
Professional
Posts: 3812
Joined: 25 Sep 2006, 01:06
Location: USA

Re: Experimental AI

Post by Rman Virgil »

------------------------------------->

* Tanks, Molotov. :) It is definitely a mouthful, hehe. ;)

* Another useful tool for this project, IMO, would be:

* The +7 Balance Engine

* Seeing as how on-topic comments seem to have provoked this threads abandonment by its author (for reasons that escape me) I will quickly broach an off-topic observation.

* Your siggy, Molotov, about Vtol Bombers can only really be addressed, IMHO, by giving Vtols different Heavy Bodies than tanks. The bodies in WZ are the Power Plants and Vtol Bombers need very powerful power plants on a strong but lightweight fuselage. That is an entirely different specification than heavy tanks. Mixing the 2 is absurd and any fussing around with the data specifications in the text files will simply not cut-it.

- Regards, RV :D
.

Impact = C x (R + E + A + T + E)

Contrast
Reach
Exposure
Articulation
Trust
Echo
.