Isn't one of the key concepts behind hierarchical memory (as described in Jeff Hawkins's work, the Serre paper I have cited, Rodney Books's subsumption, etc.) exactly that is builds hierarchically upon the regularities and modularities of whatever word it is learning in, acting in, and representing? Ed Porter
-----Original Message----- From: Robert Wensman [mailto:[EMAIL PROTECTED] Sent: Saturday, October 20, 2007 7:39 AM To: agi@v2.listbox.com Subject: Re: [agi] evolution-like systems I am not exactly sure how GA(genetic algorithm) and GP (genetic programming) is defined. It seems that the concept of gene and evolution are very much interconnected, so how we define genetic algorithm and genetic programming depends on how we define evolutionary learning, which is partially a topic of this thread. But to avoid this question and still answer your question, I could say that GA and GP needs to be pretty advanced to quell the combinatorial explosion that you speak of if it is to be used in AGI. Starting a naive evolutionary process without trying to speed things up would be pointless. It could take zillion of years to get anywhere, basically repeating the biological evolution in a computer. I believe Ben Goetzel and Novamente have some interesting points in this topic. His basic point is that a lot of necessary AGI algorithms needs to be exponential and prone to combinatorial explosions, so the key issue is to interconnect a lot of different systems so they help each other to overcome their inherent drawbacks. In their case they use a system of evolutionary learning combined with probabilistic reasoning. I have thought about another way to do the same thing, although my ideas are far from thought out. My idea is that evolutionary learning to build a world model needs to utilize the modularity of reality somehow to factorize the adaptive process. If an adaptive process is factorized, it could drastically decrease the time necessary to perform it. This is a universal phenomenon and true regardless of adaptive/learning/evolutionary algorithm. For example, the most simple adaptive process just creates a model at random, and tests whether it is correct. If a model is described using 32 bits, then the time for adaption would be in the order of 2^32. But if the model can be divided into two independent parts, the order of adaptation is only 2^16+2^16 = 2^17. Fortunately in our world, objects are somewhat independent of each other. I can rest decently assured that inner state and mechanics of my toaster, does not interfere with the inner state and mechanics of my microwave oven. This means I could hypothetically apply evolutionary learning on my toaster and microwave oven separately, and factorize the learning process in that way. In addition, in our world there seems to be classes of objects of similar design and function. If I understand the basics of a tree for example, I can apply this model to many more trees in a forrest. These regularities is also something that could be utilized to speed up adaptation, but maybe in a different way. So basically yes, making evolutionary learning work fast enough is what AGI is all about. But I do not feel that these methods to try to speed things up make it less of an evolution, at least not in my opinion. The reason I like the concept of evolutionary learning is that it implies some form of open endedness, similar to how we think the thoughts of an intellect can go in any direction. The words learning and adaptation has been too much used in narrow AI in over simplified contexts. I would like to direct a question to Ben Goetzel if he happen to read this. I am a fan of Novamente, and their ideas of quelling combinatorial explosions. But I wonder if they ever thought along the lines presented here, trying to factorize adaptation by using the modularity of reality. /Robert W 2007/10/20, William Pearson <[EMAIL PROTECTED]>: On 20/10/2007, Robert Wensman <[EMAIL PROTECTED]> wrote: > It seems your question stated on the meta discussion level, since that you > ask for a reason why a there are two different beliefs. > > I can only answer for myself, but to me some form of evolutionary learning > is essential to AGI. Actually, I define intelligence to be "an Eco-system of > ideas that compete for survival". The fitness of such ideas are determined > through three aspects: The trouble with the word evolution, is that it brings to mind Darwinian evolution which is rightly dismissed as slow and random. Computational selectionist systems can be Lamarckian or the programs can learn by themselves as well as being selected, so the speed limits of Darwnian evolution do not apply. The central dogma of molecular biology also does not apply. However this does mean that you have to use systems more advanced than GA or GP to avoid the criticisms of evolutionary systems being adequate for intelligence. Will Pearson ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? <http://v2.listbox.com/member/?&;> & _____ This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? <http://v2.listbox.com/member/?&; > & ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=55730493-023635
