## Re: [GP] Re: Agents and GA/GP

Expand Messages
• ... can you give an example? i m presuming that you don t mean constraints related to the motion or physical limitations of the aircraft. s.
Message 1 of 5 , Jan 24, 2007
> I found that by eliminating constraints required by
> analytical solutions and simply letting evolution do its job, the GP
> evolved a reactive control strategy that dramatically improved the
> aircraft's chances:

can you give an example? i'm presuming that you don't mean
constraints related to the motion or physical limitations of the
aircraft.

s.

____________________________________________________________________________________
Bored stiff? Loosen up...
http://games.yahoo.com/games/front
• Steve, you bring up an interesting issue. For many problems (for example, optimal control problems), if can be extremely difficult to derive a closed-form
Message 2 of 5 , Jan 26, 2007
Steve, you bring up an interesting issue. For many problems (for
example, optimal control problems), if can be extremely difficult to
derive a closed-form solution. The usual alternative is to make
assumptions that simplify the model enough to allow an analytical
solution to be developed. When I was studying the missile avoidance
problem, most of the textbook solutions I saw made unrealistic
assumptions (e.g., the missile and aircraft are traveling in a 2D
space, the missile is traveling at constant speed, the missile's
control laws are known to the aircraft, the missile's location is
precisely available to the aircraft, etc.). In addition, they did not
model uncertainty well at all, and could not incorporate the use of

On the other hand, an evolutionary system that uses a high-fidelity
simulator to evolve optimal strategies doesn't care about such
restrictions. Of course, it still needs to account for other
constraints, such as the physical model -- i.e., an aircraft that
attempts to pull 35Gs should be given a pretty low fitness value! So,
your assumption (below) is exactly correct.

To elaborate slightly: the classic 2D analytical solution to maximize
miss distance between an aircraft and a pursuing missile utilizing a
specific set of guidance parameters might have the aircraft make three
precisely timed turns. (See Zarchan 1992.) But the GP solution might
(for example) incorporate accelerations with turns to increase miss
distance. Further, the GP solution is easily extended to allow the
aircraft to utilize chaff/flares, jamming, altitude advantages, etc. --
something that cannot be said for the analytical solutions I've
seen. The result I saw was an improvement in predicted survivability
from something like 86% to 98%. I'm going to guess that there are a
fairly large number of problems for which GPs -- using accurate
simulations for fitness evaluation -- could evolve better control
strategies than those used today. Hope this helps!

----- Original Message -----
From: steve uurtamo <apoxonpoo@...>
Date: Wednesday, January 24, 2007 1:05 pm
Subject: Re: [GP] Re: Agents and GA/GP

> > I found that by eliminating constraints required by
> > analytical solutions and simply letting evolution do its job,
> the GP
> > evolved a reactive control strategy that dramatically improved
> the
> > aircraft's chances:
>
> can you give an example? i'm presuming that you don't mean
> constraints related to the motion or physical limitations of the
> aircraft.
>
> s.
>
>
>
>
>
>
>
_______________________________________________________________________
_____________
> Bored stiff? Loosen up...
> http://games.yahoo.com/games/front
>

[Non-text portions of this message have been removed]
• ... Call for participation ICML-2007 Workshop on Challenges and Applications of Grammar Induction In conjunction with the International Conference on Machine
Message 3 of 5 , Mar 30, 2007
-----------------------------------------------------------------------
Call for participation
ICML-2007 Workshop on
Challenges and Applications of Grammar Induction

In conjunction with the International Conference on Machine Learning,
Oregon State University, June 20 - June 24, 2007

-----------------------------------------------------------------------

Description

Grammar Induction (GI), also known as Grammatical Inference, is about
learning grammars from data. A well-known important application of GI
is natural language learning, but it is applicable in a much broader
sense to the problem of learning structural models from data. The data
typically consists of sequences of discrete events from various domains
(such as text, DNA fragments, primary structure of proteins, sequential
process log-files and musical scores), but can also include trees and
arbitrary graphs (such as metabolic networks and social networks).
Typical models include formal grammars (regular, context-free, context-
sensitive, . . .), and statistical models in related formalisms such as
probabilistic automata, hidden Markov models, probabilistic transducers
or conditional random fields.

The CAGI workshop aims at highlighting current challenges in grammar
induction with a special focus on applicability issues including:
- practical evaluations demonstrating the usefulness of the proposed
techniques,
- novel applications of grammar induction algorithms,
- noise resistant approaches,
- semi-supervised grammar learning,
- learning from partial sequences or streams,
- approximate induction and model optimization,
- experimental assessments illustrating the current limit(s) of the GI
field,
- practical complexity and scalability issues (alphabet size, noise
level, data sparseness, data inconsistency, . . .),
- evaluation of similarity learning algorithms from structured data
(pair-HMM learning, stochastic transducer learning, . . .).

Workshop Format

The workshop will include presentations of peer-reviewed papers. Each
such paper will be assigned 30 minutes, including 10 minutes for
minutes including the discussion. The day will be concluded with an
open panel for discussing the key lessons learned and pointing at
relevant research perspectives.

Submission Information

Prospective authors are invited to email their 8-page papers to
cagi07@... by the due date in PDF format. Formatting
instructions are given by the conference at
http://oregonstate.edu/conferences/icml2007/icml_format_2007.zip.
The workshop will not have a blind review process, and therefore
author names, affiliations, and contact information should appear in
number, and fax number. Electronic versions of the final papers will

Interested participants are also invited to submit 2-page position
papers. These will also be peer-reviewed and appear in the workshop
proceedings. If the workshop schedule allows, short presentations
at the end of the day may be possible as well.

Submit papers to: cagi07@...

Important Dates

Paper Submission May 7, 2007
Electronic Proceedings June 15, 2007
Workshop date June 24, 2007

Organizing Committee

Istvan Jonyer, Oklahoma State University, USA
Pierre Dupont, Universit� catholique de Louvain, Belgium
Tim Oates, University of Maryland Baltimore County, USA
Marc Sebban, Universit� de Saint-Etienne, France

Program Committee

Pierre Dupont (PC chair), Universit� catholique de Louvain, Belgium
Pieter Adriaans, Universiteit van Amsterdam, The Netherlands
Vasant Honavar, Iowa State University, USA
Istvan Jonyer, Oklahoma State University, USA
Laurent Miclet, Universit� de Rennes, France
Tim Oates, University of Maryland Baltimore County, USA
Rajesh Pareck, Iowa State University, USA
Yasubumi Sakakibara, Keio University, Japan
Marc Sebban, Universit� de Saint-Etienne, France
Menno van Zannen, Macquarie University, Australia
Enrique Vidal, Universidad Polit�cnica de Valencia, Spain

_________________________________________________________________
Live Search Maps � find all the local information you need, right when you
need it. http://maps.live.com/?icid=hmtag2&FORM=MGAC01
Your message has been successfully submitted and would be delivered to recipients shortly.