On 9/5/06, Fredrik Heintz <[EMAIL PROTECTED]> wrote:
> [re learning]
> I do not see the point in only allowing a single learning paradigm. I
> believe people learn by different methods, so why shouldn't an AI do
> the same? Both supervised learning using induction (which in some
> sense both NN and ILP does) and unsupervised learning by trial and
> error, like reinforcement learning are useful methods. Some of the
> most successful learning methods have been deviced by combining RL and
> NN (see for example the work of M. Riedmiller et al
> http://www.ni.uos.de/).
> [re learning]
> I do not see the point in only allowing a single learning paradigm. I
> believe people learn by different methods, so why shouldn't an AI do
> the same? Both supervised learning using induction (which in some
> sense both NN and ILP does) and unsupervised learning by trial and
> error, like reinforcement learning are useful methods. Some of the
> most successful learning methods have been deviced by combining RL and
> NN (see for example the work of M. Riedmiller et al
> http://www.ni.uos.de/).
We may have confused 2 things:
1. Given an AGI architecture, how the AGI learns things from the environment;
2. How the AGI architecture itself evolves via learning.
Notice that in my architecture I have completely ignored #2, also known as recursive self-improvement (RSI). I have given myself the easier task of building an AGI rather than a self-improving AI (SIAI).
I think even if NN+GA or NN+RL become very powerful, using them to evolve an AGI would still be too slow.
My approach can be termed "declarativism", first proposed in John McCarthy's paper "Programs with Common Sense" 1958, 1968, where the architecture itself is static, and what is changing is the knowledge.
Also my architecture has yet to be integrated with acting/planning. In the latter case, some form of RL would be applicable.
YKY
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
