I’d like to introduce myself to the members of this forum. My name is Karl Hirsch and I am the founder & CEO of a new AI software company called TinMan Systems. Our product, AI Builder, is an IDE that provides a complete tool set to design, train, simulate, and deploy artificial intelligence engines.

We are excited to offer this new software to the game development community in hopes of accelerating the speed, breadth, and scope of AI engines within all types of games.

This week, we released an introductory version of our software that is available to download *free* at:

http://tinmansystems.com/ProductOverview_Express.php

Please feel free to contact me with any questions. We are here to help you in any way possible.

Karl Hirsch
425-557-4360
Karl (at) tinmansystems.com
www.tinmansystems.com

Is it me or does your software only build NNs?

Is it me or does your software only build NNs?

Dave – Pleasure to meet you. Developed the TinMan AI Builder products on foundation of NN – but as a state machine. The AI Builder ‘Express version’ simply produces a single neural network (single state) structured to accommodate your inputs and outputs, and tuned mathematically as trained. The AI Builder ‘Pro version’ produces a limitless network of neural networks (multiple interconnected state machines), along with a combination of fuzzy logic, simple math models, etc…, The fuzzy logic puts context to numbers for the states, the visual design approach abstracts the NN and much of the tedium associated with NN and the modular approach makes adding and configuring and managing states pretty easy..

I have an example demo of an artificial brain of an NPC man (built in TinMan AI Builder Pro) that manages heart rate, blood pressure, breathing, thermoregulation, along with 10 primary and secondary needs / objectives (hunger, thirst, energy, boredom, fitness, social contact, hygeine, etc..) and would love to demonstrate that to you via web if you like. I took great interest in your contributions to GPGems / Wisdom, and your BehavMathforGameAI. The TinMan visual design framework of neural net based states was designed specifically to make managing many, many simultaneous human behaviors simple, and remaining open to other middleware. - Karl

Karl Hirsch | karl (at) tinmansystems.com | [color="#284b72"]www.tinmansystems.com

Do you have any examples of actual games made with your tool?

Do you have any examples of actual games made with your tool?

TinMan AI Builder Pro v1 was released in February. TinMan AI Builder Express v1 released this week (hence the free download). Until this week, our focus has been solely on government /defense related projects. We are now introducing our products to the gaming community. Because our AI platform (and TinMan Systems) is new to games, TinMan Systems is offering to work with the first few games publishers (using our professional version) on a no-charge basis to prove the value - we will produce the AI for the publisher (on spec) and deliver it along with a free Professional Version of the software to continue to modify and update. Could also do a GoToMeeting webinar to interested developers (open to times suggested by you or other GameDev moderator), to present technology and jointly discuss TinMan use for games.

[email="karl@tinmansystems.com"]Karl[/email] | TinMan Systems

Not that this hasn't been brought up in a thousand places in this forum and others, but do you realize that NNs are pretty much useless for behavioral AI in a game setting? They are simply a reactive pattern-matching device and, as such, most of the time are the wrong tool for the job. In essence, all they do is offer up a large weighted sum. In that sense, they are similar to utility-based systems or any other mathematical decision-modeling architecture. The difference is, in a utility-based model, I know what each variable and coefficient means and does -- in a NN, each and everyone one of those knobs is completely abstract and its purpose pretty amorphous. If I want behavior A to happen slightly more often when conditions X, Y, and Z are present, I twist the right knob(s) and run it again. If you want the same result, you have to pretty much retrain your NN (a potentially time-consuming ordeal) and hope for the best.

I realize that you may have deployed them in government/defense-related projects, but those spaces are increasingly looking to game technology to improve their generally tepid performance. Unfortunately, since academia and government tend to have this neurotic incestuous relationship, this adaptation of more advantageous methods has been slow in coming. That said, it seems to be unfortunate timing to be rolling out a platform based on technology that has largely been rejected by not only industry but, increasingly, by academia as well.

Can NNs do behavior? Sure. Are they useful in a production setting? Not really. They're pretty much relegated to more of an interesting "looking what this can do" sideshow.