Ben,
Thanks for your reply, It was helpful. Your answer causes me to ask in what brain-like thinking processes would MOSES be a win over just having the hypergraph itself compute candidate solutions? Hofstader's Copycat has shown that: (a) various relaxations of a given multiple constraint relaxation problem can give very different answers, (b) purtibations of various degrees can be created from states of interest in a solution space created by such relaxation, and (c) such relaxations all have an inherent fitness function corresponding the energy (or fit) of the relaxation, which fitness function can be applied to virtually any problem solved with such relaxations. Thus it would appear controlled parallel purtibation within the hypergraph could have all the potential of a genetic algrorithm, --- it would appear to be more generally applicable for most human like reasoning --- and it would interface more seamlessly with the overall Novamente architecture. I assume there are many problem, particularly where efficient fitness functions can be found, where MOSES would greatly outperform such Copycat-like relaxations. But it seems to me that in many AGI problems --- at least those based on deriving solutions, or proposed solutions based on grounded experiential learning --- Copycat-like relaxations in the hypergraph would often greatly outperform MOSES. As I said, the Copycat-like approach would have a built in, universal, fitness function. The fitness function you proposed (i.e., the strength and confidence of "Context & Procedure ==> Goal" implication) would require a relatively large sample size to be calculated with any accuracy, and each calculation would have to be made using one of the following fitness measuring methods you proposed: ===========" -- by direct evaluation, i.e. by executing P1 and seeing what happens -- by simulation, i.e. by executing P1 in some internal simulation world rather than in the real world, and seeing what happens -- by inference (PLN) "=========== Since presumably MOSES would be a relatively blind trial and error process (which would require many trial candidates solutions and their fitness evaluation for a problem before it would even have any probabilistic distributions to guide its spawning of new candidate solutions) it would seem that thousands to millions of candidates would have to be evaluated by whatever fitness function it used, often at considerable computational expense. (One would assume that a much higher percent of solutions derived by Copycat-like relaxation within the hypergraph would be reasonable (provided the system had enough relevant learned knowledge), since they would take into account much more of the relevant learned knowledge, including episodic, generalizational, analogical, and deeper semantic reasoning, and thus the number of proposed solutions that would have to be evaluated once generated would be much less.) With regard to your first proposed fitness evaluation method --- direct evaluation --- when applied to thousands or millions of trial solutions, would often not be feasible. Both because of the time required for acting in the world and because of the considerable amount of AGI computation that would often be required to perceive and evaluate the feedback from such a trial solution. Your second proposed fitness evaluation method --- Simulation --- in many cases might be more efficient. On really powerful supercomputers, thousands or millions of trial evaluations via action in a simulated world might be viable, but it would still be computationally expensive, and it would often seem that perceiving the feedback in such a simulated world using perceptual patterns learned by experience within the simulated would be computationally expensive, unless the desired goal state was very easy to evaluate, or unless their had been a hand programmed hack to speed things up. Your third proposed fitness evaluation method, PLN inferencing would often not be that much more efficient than merely attempting to solve the problem by Copycat-like relaxation, since such PLN inferencing, itself, in any case of reasonable complexity would require probabilistic constraint relaxation for the evaluation of each of the many solutions spawned by MOSES. Thus --- except in problems where MOSES's many candidate solutions could be evaluated with much less total computation than would be required to produce and evaluate the energy of the presumably smaller number of generally better solutions proposed by a Copycat-like relaxation method --- it would seem that such direct relaxations would be more efficient. But MOSES might be much more beneficial in exploring spaces for which the system has no relevant experiences to guide a Copycat-like hypergraph relaxation, or spaces, which because of their mathematical or logical nature --- such as the space of computer programs --- may be more methodically or efficiently explored by programming than hypergraph relaxation. In addition, MOSES could out perform hypergraph relaxation in the speed and degree of difference with which it could generate novel patterns. Thus, I can certainly imagine how MOSES could be used to supercharge a hypergraph's ability to solve certain types of exploratory problems, particularly ones humans are not particularly good at. But it is not clear to me that it is a win for a majority of the types of problems which the human brain performs relatively well. I would be interested in your thoughts (and those of any others on this list) concerning the above. Ed Porter -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Ben Goertzel Sent: Tuesday, December 16, 2008 7:53 PM To: [email protected] Cc: [email protected] Subject: [OpenCog] Re: What is the role of MOSES in Novamente and Open Cog?-----was---- internship opportunity at Google (Mountain View, CA) Ed, Consider a probabilistic implication of the general form Context & Procedure ==> Goal meaning (if Context C is present) & (Procedure P is executed) ==> (Goal G is satisfied) Suppose that C and G are known but P is not known Then, MOSES may be used to find P That is, if the system knows what situation it is in (C) and knows what goal it wants to achieve (G), then MOSES may be used to help find a procedure to fulfill the goal in the context Note that the goal may be derived via inferential subgoaling from other goals. Note that determining the right way to represent the current situation as a predicate (a context) is a hard inference/concept-creation problem too The fitness function may be, for instance, the strength * confidence of the implication link denoted ==> above. The fitness of a candidate procedure P1 may be evaluated either -- by direct evaluation, i.e. by executing P1 and seeing what happens -- by simulation, i.e. by executing P1 in some internal simulation world rather than in the real world, and seeing what happens -- by inference (PLN) All that is said plainly in the OpenCogPrime wikibook, but maybe the above explanation will be helpful due to its compactness? ben On Tue, Dec 16, 2008 at 7:23 PM, Ed Porter <[email protected]> wrote: Moshe and Ben, I feel I understand much of Novamente, even parts that I haven't heard explained, because it is quite similar to ideas I had developed before ever hearing about Novamente. But I have never quite understood the role of MOSES in Novamene. I do not question the power of genetic programming. Ever since I attended a 1999 lecture by Koza on what he had managed to accomplish with genetic programming --- and then spent about half an hour talking with him after the lecture and at other times at the supercomputing conference at which it occurred --- I have been very aware of GP's potential power. I am quite certain he said that in roughly a tera-opp (which could be done in one thousandth of a second on the current fastest computer) he claimed it derived a band-pass filter it took humanity almost 3 decades (until the 1940s) to develop after appearance of the earliest band-pass filters. But I haven't been able to figure out exactly how MOSES is used in the NOVAMENTE environment. For example, Koza said a key to the success of his genetic programming was having a task for which there was an appropriate fitness function. He said that in his experiments using GP to design new electronic circuitry to operate as relatively optimal band-pass filters the fitness function to determine how well each proposed solution functioned was the electronic simulation software called SPICE. He estimated roughly 99% of the his network's compute time (i.e., the above mentioned 1 TeraOpps) was spent evaluating this fitness function. >From a recent re-reading all the portion of a January 2007 version of Ben's Novamente book (an earlier version of the Open Cog documentation) that related to MOSES, I don't remember any clear explanation of what fitness function would be used to evaluate the presumably many thousands, or millions, of combo programs that would be generated in an attempt to solve a single problem. For example, in a pet brain, the pet presumably would not get a chance to try out each of the thousands of individual combo programs on a human user, to see which received a proper feedback from the user, without thoroughly exhausting the human user. (?1) So what fitness function would be used to select combo programs or direct the probabilistic distributions that are used to tune their spawning? (If this is knowledge you intend to be in the public domain.) Also, I don't understand the relationship of combo programs to the hypergraph. Combo uses a functional language which presumably seeks to do away with, or greatly restrict, side effects (obvioulsly a plus, if you are somewhat blindly cutting and pasting program fragments together) --- whereas it seems spreading activation in a hypergraph is largely all about side effects. (The 1000 to 10K synapses per neuron, plus electromagnetic field effects caused by neurons in the brain, sure sound like side effects up the wazoo to me.) I understand (a) that a combo program could be associated with individual nodes and be computed when they are activated, (b) that hypergraph nodes or edge values can be variables in a combo expression, (c) that some subset of hypergraph spreading-activation inferencing ( I didn't understand exactly which) can be used as functions in combo expressions, and (d) that the hypergraph can be used to record and generalize information about a combo program to appropriately enable inference in the hypergraph to appropriately active combo programs associated with particular hypergraph nodes. (?2) Am I correct in understanding that item (d) just listed could be quite important as a general concept in AGI learning how to automatically program, because it would allowed the non-combo aspects of Novamente to model and use probabilistic inference and attention focusing to reason about combo programs and when they should be used, combined, fed what input, or perhaps even be modified? (?3) Am I also correct in guessing that (b) and (c) would seem to enable combo programs, to, in effect, create and try out (provided a proper fitness function) novel hypergraph nodes, which would function in a manner similar to non-combo nodes largely through spreading activation? (?4) Other than what is explained above, how are combo programs and hypergraphs synergistically used in Novamente. Moshe and Ben, you are both very bright --- and you both place a lot of importance on incorporating MOSES into Novamente --- so I assume there is something important I am missing. I would appreciate it very much if you could tell me what it is that I am missing. Ed Porter -----Original Message----- From: Moshe Looks [mailto:[email protected]] Sent: Monday, December 15, 2008 1:33 PM To: [email protected] Subject: [agi] internship opportunity at Google (Mountain View, CA) Hi, I am seeking an intern to work on the open-source probabilistic learning of programs project over Summer 2009 at Google in Mountain View, CA. Probabilistic learning of programs (plop) is a Common Lisp framework for experimenting with meta-optimizing semantic evolutionary search (MOSES) and related approaches to learning with probability distributions over program spaces. Possible research topics to focus on include: * Learning procedural abstractions * Adapting estimation-of-distribution algorithms to program evolution * Applying plop to various interesting data sets * Adapting plop to do natural language processing or image processing * Better mechanisms for exploiting background knowledge in program evolution This position is open to all students currently pursuing a BS, MS or PhD in computer science or a related technical field. It is probably better-suited to a grad student, but I'm open to considering an advanced undergrad as well. The only hard and fast requirements for consideration are a strong programming background (any language(s)) and some experience in AI and/or machine learning. Some pluses: * Functional programming experience (esp. Lisp, but ML, Haskell, or even the functional style of C++ count too) * Experience with evolutionary computation or stochastic local search (esp. estimation-of-distribution algorithms and/or genetic programming) * Open-source contributor More info on plop at http://code.google.com/p/plop/, more info on the Google internship program at: http://www.google.com/jobs/students Please contact me directly (off-list) if you are interested. Thanks! Moshe Looks P.S. Disclaimer: I can't promise anyone an internship, you have to go through the standard Google application & interview process for interns, yada yada ... ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/? <https://www.listbox.com/member/?&; 5> & Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [email protected] "I intend to live forever, or die trying." -- Groucho Marx --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "OpenCog General & Scientific Discussion List" group. To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] For more options, visit this group at http://groups.google.com/group/opencog?hl=en -~----------~----~----~----~------~----~------~--~--- ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
