Table of Contents
Optimal Ordered Problem SolverOOPS: TR IDSIA-12-02; arXiv:cs.AI/0207097, July 2002 (also NIPS 2002) Slides in www.idsia.ch/~juergen/oops2002/
2003: Kolmogorov’s 100th birthday
Main result (informally)
First: Non-Incremental Universal Search
Bias-Optimality
Incremental Search?
OOPS vs Lsearch
More precisely:Searching a solver for all tasks in the sequence:
PPT Slide
PPT Slide
Recursive Backtracking in Program Space
Realistic OOPS applications
Pilot implementation: Universal FORTH-like language(many other languages / primitives are possible)
PPT Slide
Experiment 1: context free language 1n2n
Experiment 2: Towers of Hanoi
Towers of Hanoi: incremental solutions
What the found Hanoi solver does:
OOPS for virtual robots (Viktor Zhumatiy, IDSIA)
OOPS for snake robot (movie)
No Free Lunch? (skip?)
Physical Limitations of OOPS (skip?)
OOPS for Reinforcement Learning (skip?)
I could stop the talk here, but if there is time: IDSIA primitives that may help to bias OOPS
Potential OOPS primitive: Genetic Programming
Potential OOPS primitive: Adaptive Grids Milano, Koumoutsakos (ETHZ), Schmidhuber, ICANN 2001
AG Example 2 (movie)
Adaptive Grid Application: Evolving Active Flow ControlM. Milano, P. Koumoutsakos (ETHZ), J. Schmidhuber
PPT Slide
Excellent drag reduction: ~60% (movie)
PPT Slide
PPT Slide
PPT Slide
PPT Slide
PPT Slide
OOPS primitives for recurrent neural networks (Matteo Gagliolo, IDSIA, ongoing)
BTW, our LSTM Recurrent Networks are in a league by themselves
Leslie Pack Kaelbling, co-director of MIT AI lab
PPT Slide
|
Author: J. Schmidhuber
Email: juergen@idsia.ch
Home Page: http://www.idsia.ch/~juergen/oops.html
|