Hi,
I have a general question for those (such as Novamente) working on AGI systems that use genetic algorithms as part of their search strategy. A GA researcher recently explained to me some of his experiments in embedding prior knowledge into systems. For example, when attempting to automate the discovery of models of a mechanical system, they tried adding some "textbook models" to the set of genetic operators. The results weren't good - the prior knowledge worked too well, causing the GA to converge too fast onto the prior knowledge. so fast that there wasn't time for the GA to build up sufficient diversity and quality in other solutions that might have helped get out of the local maxima. The message seemed to be that prior knowledge is too powerful - it can 'blind' a search - and that if you must use it, you'd have to very very aggressively artificially deflate the fitness of instances that use prior knowledge (and this is tricky to get right). This struck me as relevant to GA-based AGIs that continually build on and improve a knowledge-base. Once an AGI learns very simple initial models of the world, if it then tries to evolve deeper knowledge about more difficult problems (but, in the context of its prior learning), then its initial models may prove to be too good: forcing the GA to converge on poor local maxima that represent only minor variations on the initial models it learnt in its earliest days. Does this issue actually crop up in GA-based AGI work? If so, how did you get around it? If not, would you have any comments about what makes AGI special so that this doesn't happen? -Ben ------------------------------------------- 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=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
