Interactive Evolutionary Computation
Interactive Evolutionary Computation (IEC) is an umbrella term covering a range of techniques where the user interacts with the Evolutionary Computation search process to direct or modify the search based upon subjective preferences.
In a review of IEC, Takagi (2001) presents two definitions corresponding to a narrow and a broad view: In the narrow view IEC is seen as “the technology that EC optimises the target system based upon subjective human evaluation as fitness values for system outputs” (Takagi, 2001, p. 1275). This narrow view is more traditionally referred to as Interactive Genetic Algorithms (IGA). On the other hand, the broad definition is seen as “the technology that EC optimises the target system having an interactive human-machine interface” (Takagi, 2001, p. 1275), where the user modifies the GA parameters. A similar, but slightly broader, categorisation of IEC is also made by Parmee (2002).
In the IGA the user effectively replaces the fitness function. The classical example of this is the Biomorhps of Dawkins (1986), in this instance the user is presented with images of a series of candidate designs and selects a subset for reproduction. Their usage is normally in situations where it is not possible to provide a quantifiable fitness function, i. e. in the case of evaluating designing for their aesthetic content.
The approach adopted by IGA’s has been successfully applied across a range of domains: Graphic art (e. g. Unemi, 2000; Todd and Latham, 1999); Industrial design (e. g. Graf, 1995); Face image generation (e. g. Takagi and Kishi, 1999); Speech processing (e. g. Watanabe and Takagi, 1995); Geological modelling (Wijns et al., 2003). The reader is referred to Takagi (2001) and Banzhaf (1997) for a more comprehensive review of applications.
While the IGA has only found limited usage in conceptual building design (Buelow 2002), it has been argued that such systems have a major role during conceptual design and “can fit naturally into human design thinking and industrial design practice” (Eckert et al., 1999).
Despite the obvious success of this approach it does suffer shortcomings, one of these being human fatigue. In a typical EC search the population size is usually one hundred and may evolve for a considerable number of generations. Human evaluation of this large number of individuals leads to not only inevitable fatigue, but also it is cognitively difficult for humans to rank or choose between this number of individual concepts generated by the system.
One approach to this problem is to use smaller population sizes across fewer generations. On more complex design problems with high dimensional search spaces a small population size may not cover the search space adequately but also exhibits rapid convergence.
To overcome the small population size issue, predictive evaluation techniques can be adopted. Here the user assigns a fitness score to a few selected individuals and the IGA predicts the fitness of the remaining individuals. This prediction can be either based on machine learning techniques (i. e. Neural Networks, Biles et al. 1996), or a Euclidian based similarity measure. The difficulty with Euclidian techniques is to develop a scale which mirrors the human evaluation scale.
The IGA may only have a limited role to play as an exploratory design tool in its fullest sense. If EC approaches are used as exploratory tools then it is necessary to allow the user to explore, or to move the search to, differing areas of the solution space which the user (i. e. expert designer) finds particularly interesting or due to a different emphasis on particular design requirements, and to have the ability to return to those areas for later investigation.
Currently it would seem that while IGA’s are very effective in allowing a user to guide the search through the solution spaces, they do not readily allow the re-visiting of particular choice points. This ability to move the search, or focus at a specific region, is particularly important during conceptual design where the actual requirements themselves are being re-defined through the exploration process.
When the broader definition of Takagi (2001) is considered, again variants can be identified. One approach which uses human interaction in its fullest sense is the Human Based Genetic Algorithm (HBGA) of Kosorukoff (2001) and Kosorukoff and Goldberg (2002). The HBGA extends the IGA approach of the user driven selection, by allowing the user the option to perform and modify the GA operations of crossover and mutation. With the motivation that humans prefer to be creators rather than critics, in essence the evolutionary search process is now guided by human innovation. Moreover the HBGA is seen as a key member in collaborative problem solving and is playing a major role in the Free Knowledge Exchange (FKE) project and the DISCUS1 project of Goldberg et al. (2003).
An alternative, and more traditional, approach is where the user is allowed to set or to modify GA parameters. Thus the work of Parmee et al. (2000) allows the user to set design objective preferences at the start of the GA search process, whilst the work of Mathews and Rafiq (Mathews, 2000; Rafiq et al., 2003; Rafiq et al., 2001) allows user modification of GA parameters during the GA run.