Parallel and Distributed Tools for Evolutionary Computations
------------------------------------------------------------

Over the past 20 years, because of their generic solving 
capabilities, evolutionary algorithms (EA) have become more 
and more popular. They have demonstrated their usefulness at 
solving real world problems in almost all of the hardest problem 
areas. They may not be the most efficient methods for all problems, 
but they are always applicable and, in many of the hard cases 
where optimal solutions are intractable, they represent one of the 
better alternative. The price to pay for this genericity, however, 
is high (sometimes huge) computational effort. Thus the need for 
parallel and distributed evolutionary computation tools.

This seminar is structured into three 2-hour parts. Part I is about 
fundamentals, part II about existing tools, and part III about a new 
software architecture that we are currently developing for distributed 
EA deployment. Our main objective is to survey the field of evolutionary 
computations (EC) in order to identify potential areas of parallelism. 
We will not be introducing any new EA per se, nor describe bleeding 
edge techniques, our aim is more to highlight the range of existing 
EA flavours and their embarrassingly parallel nature. We will not 
introduce any new parallel paradigms either, what we wish to achieve 
is to identify what can be parallelized and how to implement this 
parallelization as simply and efficiently as possible on a large 
supercomputer.

For part I, we will thus start by justifying the need for EC, and 
then carry on with a survey of its main flavours: genetic algorithms, 
evolutionary strategies, evolutionary programming, and genetic programming. 
We will also address several important paradigms in EC, like island 
models, coevolution, hybrid methods, and multicriteria optimization. 
Finally, we will also overview some fundamental concepts of parallelism 
like distributed memory and message passing.

In part II, will focus on hardware and software tools. In particular, 
we will first introduce the CLUMEQ consortium for high performance 
computing and its new Colossus supercomputer. Then, we will focus 
on requirements for building software tools that enable rapid prototyping 
of evolutionary systems on such parallel computers. Finally, we will 
overview Open BEAGLE, one of the fairly well-known and freely available 
framework for distributed EC, and discuss some of the weaknesses that 
it shares with other frameworks: general complexity and lack of documentation. 

Finally, part III will present an on-going project called DEAP that 
we hope will constitute the foundations for the next generation of 
EC applications, to enable the rapid and efficient development of 
distributed evolutionary algorithms, thus facilitating and accelerating 
the validation of new research ideas.