LEGG, Shane

<p>Oct 21, 2012 11:14 PM

see his YouTubes...

def  Intellig measures an agent's ability to <b>achieve goals in a wide range of enviros</b>.

Bingham... ability of agent to solve new problems...

Colvin... LEARING to adjust to new enviros.

ability to act purposefully , think rationally, and deal effectively with the enviro

LEGG emphasizes achievement of goals...  in a wide range of enviros...

broad range of tasks...

an agent's ability to achieve goals in a wide range of enviros.

cannot ignore computational cost...

prior to Hutter nobody had even formally def. comput cost...

1 3 5 7...  what comes next...  simplest explanation uses Ock razor to get 2-1...

use Ockham's razor...

Bayes Theorem...  updating prob hypotheses...

simplex explanations are more likely...

Kolmogorov complexity.... computatble description  eg

1 with a trillion zeroes after it....   there's the descr... much shorter than the actual thing...

same with pi...

Ockhams razor  ties in with Kolmogorov docmpl...

length of a program...

probability of a program  decreases with its length by 1/2...

Ockham... simpler hypotheses are given higher prob than complex ones...

he introduces Kolmogovorv in 4/12 at minute 4....

predict  sequence given what you've seen...

chsi... is a distribution... Solomonoff induction...

we don't know what's generating the seq.

what is the prob. that the next digit is ... 0 etc.

works ridic well....

for any unkonw computatble distrib.

get the expected total prediction error...

compute this over all possibilities...

Kolmogorov depends on the language that is expressed in...

apply to QM.. easy to write down (Schrodinger) but v hard or intractible to compute...

how do u approximate  Kolmogorov or Solomoff induc... must apporximate it...

5/12  RL = REINFORCEMENT LEARNING

agent interacts with enviro  gets rewards...  when  rewarded that action is more probable..

tries to compue next action  (has total hx of all states, actions and rewards)

Agent's goal is to max reward....

 </p>