Marcus Hutter
<p>May 25, 2011 09:20 PM
 
    Marcus Hutter Universal AI  now at Canberra Univ.
 
goal:  construct a single universal agent that learns to act optimally in any environment.
 
reduces  AI to a "pure" comput problem
 
What is AI"  creativity, reasoning, assoc, generalization, pattern recog, problem solving,
memorization, planning, achieving goals, learning, optimiz, self-preservation, vision, langu processing,
classif, induct, deduct... etc.
 
What is AI?  collection of 70+ defs of intelligence  at
www.vetta.org/definitions-of-intelligence/
 
real world is nasty:  uncertain, partly unobserable  etc.
 
Intelligence measures an agent's ability to achieve goals in a wide range of environments
S. Legg and M. Hutter.
 
Induction -> prediction -> decision -> action
 
induction:  find model of world economy
prediction:  use model to predict future stock mkt
decision:  whether to invest in stocks or bonds
actions:   trading large quantities of stocks influences the mkt.
 
Science:  Induction = Occam's Razor
H1  all emeralds are green   simplest hypothesis
 
info theory and kolmogorov complexity
 
shortest description of object is best explanation  (eg smallest Turing program);
shortest program for a string on a Turing machine T;
 
Bayesian prb theory:  goal is to get posterior prob  P(Hi given D) of Hi after having seen data D.
 
Epicurus:  if more than one theory is consistent with the observations, then keep all thoeries.
 
how to deal with this....
 
assign  high prob to simple hypotheses in inverse proportion to (log of) length.
 
Ray Solomonoff:  died last year....  born in 1926...
 
Solomonoff. combined Occam, Epicurus, Bayes, etc.
 
seq decn theory:  minimize expected loss (or max expected gain)
 
greedy maximizaiton:   short-term max  (take pawn with Q then lose Q);
NO... must maximize  gain over life-time of agent.
choice of horizon.
 
AIXI  model in one line:  CLIMAX... most intellig agent in the universe in one line!!!!
 
action, reward, observation,  Universal TM= turing machine
 
k = number of cycles  a  = action kth;   r = observations
 
must take expectation with respect to unknown universal distribution.
 
weighted average of all models  (complex models get small prob; simple models get large prob)
must sum overunobserved future probabilities.
 
agent performs action, makes observations,  gets rewards or sustains losses;  reformulates actions;
 
implementation difficult;  its not computable;
 
Levin search:  Jurgen Schmidhuber  did Levin search... did implement on a normal pc;
maze, towers of hanoi, robotics;
 
compress 100mb of wikipedia... related to intelligence;
 
criticisms:  AIXI  is obviously right or obviously wrong...
No-free-lunch  theorem....  means cannot estiamte a random function
(but world HAS structure);
 
bridge the gap btn Universal AI theory anbd AI practice... go beyond tic tac toe
better to raise  AIXI (pron IKE see) to believe in hell ... so it continues to live.
 
self-maintenance  (battery life, oil for a robot)...
 
must measure intelligence relative to a certain goal...   what are the worthwhile goals;
intelligence is only relevant to goals  MUST specify goals...
we want a system that  seeks or solves human specified goals;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 </p>