On Nov 10, 2007 11:42 PM, Edward W. Porter <[EMAIL PROTECTED]> wrote:
>
> You say there is no magic in AIXI.  Is it just make believe "Let X be the
> best way to solve problems. Use X", or does it say something of value to
> those like me who want to see real AGI's built?
>
Some observations that come to me from reading Marcus Hutter, you can
see their worth:

(1) he gives rates of convergence for a posterior distribution to a
real distribution, assuming that the real distribution has a non-zero
prior probability

(2) he shows that a weighted (=mean taking) predictor converges
possibly exponentially faster than maximum likelihood predictor

(3) he shows that the expectimax algorithm is optimal given unbound
resources = you can view things "functionally": best possible policy
(program, strategy), or iteratively (expectimax)

(4) he discusses the issue of choosing horizon, reinforcement learning
(=RL) work usually uses geometric discounting, Hutter shows that it
gives (ln 0.5 / ln d) effective horizon (where d is the discount
rate), (and it would be "theoretically" justified when the agent has
probability d of surviving to next cycle)

(5) he discusses RL from a general stance, e.g. classes of
environments, application to learning frameworks more specific than RL
(supervised learning, optimization) (but only theoretically)

(6) he discusses the issues of dividing computation resources into
using currently best strategy and searching for new strategy (in his
"time-optimal algorithm for all well-specified problems")

(7) his computational AIXItl model, assuming that it's the best out
there since Marcus didn't come with something better :-), justifies
some practical approaches of letting the competing policies estimate
their expected utility: AIXItl only allows policies for which it can
find a proof that the policy doesn't overestimate its utility, Eric
Baum's market economy uses the policies' claims as a currency
(cheating policies go bankrupt), accuracy-based Michigan-style
learning classifier systems XCS use (evolutionary) selection pressure
on the accuracy of the policies' claims

(8) the "Kolmogorov-complexity-inspired distribution over programs" is
related to new "better than genetic programming" approaches
(Schmidhuber's OOPS, MOSES) (but perhaps only distantly)

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=8660244&id_secret=63908199-d1781a

Reply via email to