This is a significant improvement - good work.
Only having skimmed, my first objection is that you set the initial
bar for a self improving program too high. IMHO, the appropriate
initial program _is_ a burbling infant. You don't, for instance, write
from scratch a data structure for objects: You hook a sufficiently
powerful learning system up to raw pixel data and let it develop it's
own representations. You don't, initially, provide any connection
between a control line for an actuator and the sensors housed in a
limb: you allow the system to discover a causal correspondence
itself. You allow it to discover cause and effect, object permanence,
the distinction between internal and external stimuli, planning,
etc. If you can construct a machine that does that then the problem of
general intelligence is solved, and self improvement at the source
code level will be a historical footnote.
The hypothesis: It is possible, at least in principle, to construct a
machine which acquires all the skills and knowledge it requires to
function in the world via learning. Human provided domain-specific
knowledge, even of highly abstract or primitive domains, is
unnecessary. The question then is what is the minimal such machine,
and how do we go about building it.
This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:11:24 MDT