The abstract is funny. "A new theory... known as A-Design is introduced..." How can this article be introducing something that is new and known?

      The authors note that objectives and constraints are usually required to be fixed prior to design. While A-Design is interesting in the way it does not have this requirement, systems before A-Design have focused on dynamically adding goals and constraints. A-Design is different in that the user can change preferences thereby changing the problem to be solved slightly during the design process. The other systems add or discover new goals and constraints while solving the same problem throughout the design process.

      A list of the major differences between A-Design and previous work is given on page 580. It seems that only the last item mentioned is accurate. Iteratively improving upon a design to meet objectives and rich representation have been common themes throughout previous design systems. Keeping possible design alternatives and the method that manages them is what truly sets A-Design apart.

      The justification given for calling A-Design's agents animate has no relevance. How can feedback that allows changes in focus and improve search make it obvious that the agents are full of life (or any other valid definition of animate)? So far the terminology used is very confusing. A-Design is referred to as a methodology, a strategy, a theory, a technique, a process, and having sub-systems to its theory. Needless repetition is also prevalent, many sentences repeating terms and rewording what has already been said in immediate context. I'd give examples, but there are over 25 to choose from in the first few pages.

      The authors feel it necessary to point out that the theory requires that the representation and agents that work with the representation be defined (pp. 581). How would it work if there was no way to represent the problem or design space?

      The discarding of designs evaluated as poor can lead to the local maximum problem seen in other algorithms such as hill climbing. There is no mention of this possible limitation despite it being a well known issue in genetic algorithms. A-Design further adds to the local max. problem by giving preference to the agents that produce the pareto optimal and good designs.

      When speaking of past design alternatives it is unclear if these consist of only the past pareto optimal designs or also include the good or poor designs. Judging from the diagrams, I'd guess just the pareto optimal. There is a reference to Figure 6 (pp. 588), but I couldn't find it. How disappointing not to see a gear embodiment.

      The author's tout A-Design as maintaining populations of both designs and agents. This doesn't seem very noteworthy since the agents themselves don't change. Only the list of agents that have the most influence on the design population changes. The evaluation of the designs should guide the characteristics of the design population. This would naturally result in more designs by some agents than others. By directly manipulating the agents that are active the whole genetic evolution process as a search strategy gets by-passed by prematurely limiting the search space in a manner that can lead to stagnation.