Markus Aschinger, Conrad Drescher, Gerhard Friedrich, Georg Gottlob, Peter Jeavons, Anna Ryabokon, and Evgenij Thorstensen. Optimization methods for the partner units
More than 25 years ago together with Siemens we started to investigate the possibility of substituting the classical rule-based con guration approach by model-based techniques. It turned out that in those days only constrained programming (CP) had any real chance of meeting the application demands. By exploiting CP we were able to signi cantly improve the productivity of highly trained employees (by more than 300%) and to substantially reduce software development and maintenance costs (by more than 80%) (FFH+98). Consequently, CP has been our method of choice for problem solving in industrial projects since 1989.
Some years ago, we started to investigate answer set programming (ASP) techniques (BET11), mainly because of the possibility to apply a very expressive logical rst-order language for specifying problems. It emerged that, by using simply problem encoding, we were able to solve di cult real world problem instances witnessing the enormous improvements of logic programming over the last decades (ADF+11).
Although ASP and CP have proven their practical applicability, we will point out challenges of large problems of the electronic and the semiconductor industry. In particular, we will stress the power of problem-speci c heuristics (GKR+13; STW+13) which turned out to be the key in many applications of problem solvers.
Looking at the famous equation \algorithm = logic + control" (Kow79) most of the current work in the AI community assumes that control should be problem-independent and only the logical speci cation depends on the problem to be solved, i.e. \algorithm = logic(problem) + control". It is not surprising that for the current problem solving technology this is a practical approach up to a certain size of the problem instances, since we deal with NP-hard problems in many cases. However, it is observed (and examples are given (TFF12; MF15)) that problem-speci c heuristics allow enormous run-time improvements. This success is based on problem-speci c control, i.e. \algorithm = logic(problem) + control(problem)". Unfortunately, the design of such problem-speci c heuristics is very time-consuming and redesigns are frequently required because of recurrent changes of the problem. Interestingly, humans are very successful at developing such problem-speci c heuristics. Therefore, we argue that the automation of generating problem-speci c heuristics with satisfying quality is still an important basic AI research goal with high practical impact that should be achievable (Pea83).