








Jürgen Schmidhuber's
course
MACHINE LEARNING & OPTIMIZATION II



General Overview.
This is a followup course (Vertiefungsfach)
for those familiar
with the material of
Machine Learning & BioInspired Optimization I,
focusing on learning agents
interacting with an initially unknown world.
From foundations of algorithmic information
theory to asymptotically optimal yet
infeasible methods showing the ultimate limits of machine learning,
and all the way down to practically useful tricks for recurrent
nets.
The preliminary outline below does not reflect the
order of topics.
 
Course material.
We will often use the blackboard and ppt presentations.
In the column below you will find links to supporting material.
Don't worry; you won't have to learn all of this!
During the lectures we will explain
what's really relevant for the oral exams at the end of the semester.
But of course students are encouraged to read more than that!
Thanks to
Marcus Hutter
for some of the material.




More on Recurrent Neural Networks.
RNNs can implement complex algorithms,
as opposed to the reactive
input / output mappings of feedforward nets and SVMs.
We deepen the discussion of ML/BIO I on
gradientbased and evolutionary learning algorithms
for RNNs.
 



Advanced Reinforcement Learning.
We discuss methods for learning to maximize reward
with nonreactive policies in
realistic environments, where an agent needs
to memorize previous inputs to act optimally.
We also discuss ways of implementing artificial curiosity
for improving some robot's world model.
 



Algorithmic Information Theory / Kolmogorov Complexity.
We discuss fundamental
notions of information, compressibility, regularity, pattern
similarity, etc. This theory provides a basis for
optimal universal learning machines.
 



Optimal universal learners.
Bayesian learning
algorithms for predicting and maximizing utility
in arbitrary unknown environments, assuming only
that the environment follows computable
probabilistic laws.
These methods are optimal if we ignore computation time.
They formalize Occam's Razor (simple solutions preferred).
 



Optimal universal search techniques.
We discuss
general search techniques that are optimally efficient
in various senses, including
Levin's universal search, Hutter's fastest algorithm for all
welldefined problems, and Gödel machines.
 

