January 28, 2015
Attending:
Thom
Bill
Karthik
Nic
Lee
Mike
Wren
Frode
Agenda:
Gecco Papers -3
Thom, Bill, Nic
Lee-
Asymmetry on parents
Collaboration with David Clark on math-finite terms and finite algebra, Clark has algorithms that are not evolutionary
Pucks; bonding, will then be able to do push based
Nic-
Crossover revised paper; 2 problems, crossover bias and tournaments
Lee and Nic-
Measurement of success in systems, best approximation, best fitness
Both metrics are potentially useful, and when you don’t know…
Thom and Nic-
Generalization or how well matches the training set-over fitting?
No validation, but spot-check via graphs; stable, solid approximations over test area
Adding crossover bias-fitness becomes larger
Bill-
Generalization and how to promote it in GP runs
Lee, Thom and Bill-
Papers from journal; collectively read and discuss
Trade of fitness and validation of fitness
How different system generalize-but not making better at generalization
Bill-
Promoting generalization in symbolic regression
Smaller programs generally generalize better with better fitness
Random mate selection, survival step-fitness and fitness generality
Lee, Thom and Bill-
2 fitness cases,
Survival-how well fitness to 1, and how well fitness’s compare
2 sets similar-performance on one dependent on other
Size connects to post run simplification
Simplify with hill climber after
Each step turns some genes off (remove sub-expression)
Lee-
Trees; mutations with replacement of sub tree with parent tree
Preserve syntax
Lee, Bill and Thom-
Post run simplification
Instead of pure hill climbing
Turn off two genes and turn one gene back on
Higher probability of turning things off then on
Frode tried to do systematic testing; same simplification but as genetic operator and how it works with size and generalization; all a wash, simplifying during a run has all sorts of repercussions
Post run simplification for generalization
Bill, Nic and Lee-
Epigenetic paper; danger of varying multiple things and their repercussions and relations within the system
What is the epigenetics contributing? Is it just turning off your other additions?
Tough sell without solving some unsolvable problems
Clarity about base system and what is being changed
Explanation of the nuances of ideas
Rational for starting point
2 stack based systems; epigenetics is easier to do in stack-based systems
What happens when stuff gets turned off?
Thom-
Push for hill climbing; doesn’t take secondary
Keeps child if better at every test case, even with out lexicase
Wren
More LatB data
Calc concentration eq; relates to what the cell looks like
Models with Sarah
Mike
Documentation; example programs, how to put in your own data
Java version of push
Unwitting GP, Parasitic computing
Zeke’s tree based GP
Quill, Gorilla repl
Outsiders get started with out intensive tutorial
Cleaner; FourPush?
Karthik
Trying inductive synthesis with Sketch (having trouble with strings)
Running Tom’s Benchmarks
Frode
Thinking about using simplification before parent selection.
Unclear whether this would help.
Tom
GECCO Papers
Asymmetry
-In tree based GP, the root node is more important.
-Biological Reproduction is generally asymmetric
*GPTP wants “something” by 2/7.
*Gecco Workshop deadlines: 4/3