I would love to get some comments from members of this group. Watch the linked you tube video and comment away.

http://www.youtube.co...

It was quite interesting. However, I'm thinking that evolving critters in an artificial world might be a bit on the slow side for developing an AI. If you just want to evolve the intelligence aspect of the machine I would throw out the eating/procreating stuff and worry less about the mechanics (unless the AI you want is part of a robot). E.g. there are no intelligent insects on the planet because an early choice on structure limits their size. Likewise birds have limited brain power because a large brain takes too much effort to lift.

Personally I'd probably go for building a rule-based machine and let the rules evolve. I think that has been shown to work for mechanical/process control where the goals are simple and the systems fairly linear, maybe it will work as a general approach if the right kind of computing platform is used.

A question:

Is "intelligence" the same as the lack of "stupidity"?

It's an interesting demo and presentation.... However, The exchange(signal, sign, and .....) energy that define and form intelligence and life might be based on Bosanic String Theory concepts with 10-26 dimensions. That's my thinking.

Cuong Nguyen

OK, string theory certainly packs a lot of variables (attributes, dimensions) into the small end of the configuration spectrum. And this gets people's panties all in a bunch. "Magic is real!" goes the cry. "And we now have a mechanism to make magic literal and physical." But these opinions are driven by desire more than reason. When you know anything about strings, quarks, etc. etc., you know that although this world of the many-dimensions and weird interactions, is only many dimensional and weirdly interactive when the strings or quarks are isolated. Of course this translates to "theoretical" because isolation at the bottom is either fleeting or the result of the most amazingly macro machinery (huge machines that use up unbelievable amounts of structure and energy to fight the natural order of things). Another way to say this is that it takes a whole ____load of parts to make anything that can encode an abstraction and reason on that abstraction. More than that, knowledge, intelligence, reason, communication, etc. all of these necessitate stability and latency in the system from which they arise. Structural stability is not the strong suit of strings or quarks. If someone would bother explaining to me how their pet string or quantum intelligence theory satisfies the structural stability and latency requirements, or why structural stability and latency are not a fundamental requirement of intelligent systems, then I am all ears and eyes.

Kevin,

When you talk of an aspect of a virtual world you know you are talking about an abstraction, right? What "eating" and "procreating" represent in PolyWorld can be generalized to any fitness filter in any evolving system. No matter what you are attempting to evolve, some aspect of the current individuals in the population must be subject to a test that determines survival (viability and access to the future). Any pure "knowledge" system will have also to map aspects of its structure to the more general concepts of eating (acquisition and effective use of necessary resources) and procreation (which attributes are judged more viable and how to make individuals attracted to other individuals that exhibit such attributes).

Maybe PolyWorld is too literal for you. The general ideas being explored in this intuitive visualization are important in any evolving system. I am quite sure that the authors were not especially interested in advancing the art of polygon animation or the evolution of legs and arms for self propulsion. They just wanted to play with the most general aspects of evolving systems and choose a domain of focus that allowed easy human cognition of the simulation as it ran.

Randall

Randall,

To me, the "structural stability and latency requirements" will be seen in closed string....

Cuong Nguyen

If we can view intelligence as an ability to deal with available resources, regardless of how simple or complex, than we can begin to approach the problem with regard to its components. With respect to the polyworld representation, the resources were the evolved geometric figures. Briefly mentioned in the film but left out of this discussion are the system motivators. Intelligence or any other actions taken by humans or any other thing or creature will require them. One of my questions of the film was, were the list of behavior modifiers broad enough to direct the creatures toward intelligence? If their list was not complete, than what did it leave out? I have ideas regarding this for another discussion.
It often seems inevitable that any discussion of AI opens the doors of the universe to illuminate the perceptually impossibly complex path of human intellectual decomposition. Instead directing our attention toward a minimal understanding of even simple animal ability and allowing that to give direction, we take seemingly wild guesses at the unprovable and profoundly complex. Doesn’t it seem they are saying by these statements “If I can make others think that they can’t understand this stuff than they won’t know that I don’t understand it either”, and so endless pointless discussion. The problem is nobody gets anywhere. Let’s do something else.

Kevin,

When you talk of an aspect of a virtual world you know you are talking about an abstraction, right? ...

Maybe PolyWorld is too literal for you...

Randall

Yes, but it's an abstraction of the real world. I'm not sure that modeling the real world is useful in that the goal is usually just survival - it took billions of years to produce humans by that process (from the PolyWorld level), there's no guarantee a model of the process can go any quicker and/or produce results.

A problem with "human intelligence" is that in order to improve it you have to fiddle with how the entire brain is built at the genetic level - how many neurons, chemistry etc. So for humans I suspect most of the algorithms used in intelligent thought are less pre-programmed and more learned from society/experience - consider the time it takes to learn to read.

I would prefer to see a machine that can model the world and reason about it programmed in a way that can be evolved, where the goal is just to "think better". It could be endowed with a lot of useful stuff from the get-go - like the ability to communicate, knowledge of physics and chemistry etc., that humans normally have to learn.

AIs in sci-fi that worry about survival tend to be troublesome :-)

Kev.

Kevin,
I agree with your appraisal that the Polyworld model lacked evolutionary timeliness. What do we expect to see of that evolutionary process that is unavailable to us from an inspection of our own? The changes during the first 4 billion years that brought us to humanity were very subtle. Will we be able to detect their relevance? We have before us a great variety of species and knowledge of how they evolved to be what they are. Seeing the evolutionary process unfold before us will demonstrate the benefit of intelligence but not necessarily its internal processes. Indeed some form of intelligence may evolve from Polyworld one day but that will just strengthen the case for evolution. And we will still be facing an in depth analysis of the behaviors that Polyworld evolves, this is something we can accomplish now from analyzing our own behavior.

I disagree with your perspective regarding the need to fiddle with human genetics to improve intellectual performance. However, I would like to support you regarding your statement that thought may have its bases in what is learned from experience. Only by learning from its experiences can a specie have relevance regarding what exists within its environment. All of thought has a basis in an entities perception of its environment. The elements of the useful stuff you mention “ability to communicate, knowledge of physics and chemistry” relate to and rely on an understanding of a need to utilize this information. Such needs are products of other conceptual aggregations. We take small steps regarding learning what motivates us. Such databases you suggest, without any relevance to the entity, would have no meaning and not be used.

Kevin,
I agree ...

I disagree with your perspective regarding ....
We take small steps regarding learning what motivates us. Such databases you suggest, without any relevance to the entity, would have no meaning and not be used.

Part of the evolutionary process is the trade-off between hard-coding and learning. For humans the bias is towards learning and following social norms. E.g. you could genetically encode language itself, but it's cheaper genetically just to encode the ability to learn language. The downside of this is that geographically isolated groups then evolve different languages.

For a machine the cost function is different - learning is expensive, hard-coding is cheap. Monica's model-free language analysis methods are a prime example of this: human brains can never accumulate the quantity of data to apply the model-free methods reliably, a computer can and will be able to reason about the language in a very short space of time. I.e. dumping the sum total of (useful) human knowledge into a machine (or just making it directly available) is possible such that the machine does not have to learn it in the way humans do.

Another way of looking at it is that there is an energy cost with carrying data, keeping the neurons alive in a biological system is relatively expensive, machines can commit to low energy storage (flash/memristor) and also can load data from remote systems as needed - human information sharing is somewhat less efficient. E.g. if my AI finds itself among foreigners it may choose to unload its understanding of English and load up (say) Spanish - which is something to bear in mind if you are designing a general purpose AI/robot.