This is a 1-hour talk about Evolutionary Computation for optimization, written by Claus Aranha.
This talk focuses on discussing the problem of how to generate new algorithms through composition.
The talk begins with a simple explanation of Optimization through EC, then a digression into the multiplication of algorithms. Then it moves to the main point about how we need an organized way to generate new algorithms for specific tasks/domains.
This talk was first given for the Master Course on Evolutionary Algorithms at the University of Malaga, on 2022/12/15.
Evolutionary Frankenstein: The change in Evolutionary Computation
from metaphor algorithms
to component research
Evolutionary Algorithms have been effective methods for hard optimization problems since the 80ies. Their key idea, as the name indicates, is to use a process that is similar to/inspired by biological evolution to continuously select ever improving solutions to a problem. In the same period, Ant Colony Optimization (ACO) showed great success for solving graph search problems by using ideas based on how ants navigate their environment. Unfortunately, based on the success of these algorithms, the 2000s have seen a true parade of works proposing methods inspired by increasingly bizarre metaphors, such as cats, whales, music, and even zombies. In most cases, these "metaphor algorithms" differ very little from traditional meta-heuristic methods, and stake their novelties only in the metaphor itself. More recently, the field has started to move away from "creating new algorithms" to "developing components", where each component is rigorously studied for its capacity to contribute to the overall search. The search algorithm itself is now created by an automatic process, that chooses the components that better fit each target problem.
This talks offers an overview of the above trajectory, explaining Evolutionary algorithms overall, the initial motivation for creating new algorithms, the "metaphor craze" era, and the recent push for more methodical research on meta-heuristics.
TODO
- Soresen's "Metaphor Exposed"
- Bezerra's paper on Composition
- My papers on composition
- Yuri's paper on component analysis
- Jair's references (Armas, Lones, etc)
- Fully List papers on "Further Reading" (Links, etc)
- Improve third part of the presentation (less text on slides!)
- Add Yuri's work on component analysis
- Write a script
- Measure time
You are free to re-use / modify any materials in this talk that I've created. Some of the figures have not been created by me, so I recommend that you check the usage rights from the original authors (usually listed at the file name, or in the tex code).