--- On Tue, 11/18/08, Mark Waser <[EMAIL PROTECTED]> wrote: > add-rule kill-file "Matt Mahoney"
Mark, whatever happened to that friendliness-religion you caught a few months ago? Anyway, with regard to grounding, internal feedback, and volition, autobliss already has two of these three properties, and the third could be added with an insignificant effect. With respect to grounding, I assume you mean association of language symbols with nonverbal input. For example, a text-based AI could associate the symbols "red" with "rose" and "stop sign", but if it lacked vision then these symbols would not be grounded. To ground "red" it would need to be associated with red sensing pixels. In this sense, autobliss has grounded the symbols "aah" and "ouch" which make up its limited language by associating them with the reinforcement signal. Thus, it adjusts its behavior to say "ouch" less often, which is just what a human would do if the negative reinforcement signal were pain. (Also, to address Jiri Jelinek's question, it makes no conceptual difference if we swap the symbols so that "aah" represents pain. I did it this way just to make it more clear what autobliss is doing. The essential property is reinforcement learning). Also, autobliss has volition, meaning it has free will and makes decisions that increase its expected reward. Free will is implemented by the rand() function. Behaviorally, there is no distinction between free choice and random behavior. Belief in free will, which is a separate question, is implemented in humans by making random choices and then making up reasons that seem rational for making the choice we did. Monkeys do this too. http://www.world-science.net/othernews/071106_rationalize.htm Autobliss lacks internal feedback, although I don't see why this matters much. Neural networks often use lateral inhibition, activation fatigue, and weight decay as negative feedback loops to make them more stable. Autobliss has only one neuron (with 4 inputs) so lateral inhibition is not possible. However I could add weight decay by adding the following code inside the main loop: for (int i=0; i<4; ++i) mem[i] *= 0.99; This would keep the input weights from getting too large, but also cause autobliss to slowly forget its lessons. It would require occasional reinforcement to correct its mistakes. However, this effect could be made arbitrarily small by using a decay factor arbitrarily close to 1. Anyway, I don't expect this to resolve Mark's disagreement. Intuitively, everyone "knows" that autobliss doesn't really experience pain, so Mark will just keep adding conditions until nothing less than a human brain meets his requirements, all the time denying that he is making choices about what feels pain and what doesn't. -- Matt Mahoney, [EMAIL PROTECTED] ------------------------------------------- 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=120640061-aded06 Powered by Listbox: http://www.listbox.com
