Learning As Applied to Simulated Annealing

Stochastic combinatorial optimization techniques, such as simulated annealing and genetic algorithms, have become increasingly important in design automation as the size of design problems have grown and the design objectives have become increasingly complex. However, stochastic algorithms are often slow since a large number of random design perturbations are required to achieve an acceptable result— they have no built-in “intelligence”. In this paper, we show that incremental, statistical learning techniques can improve the quality of results and reduce the number of expensive cost-function evaluations for stochastic optimization for a particular solution quality. In particular, simulated annealing was selected as representative stochastic optimization approach and the cell-based layout placement problem was used to evaluate the utility of such a learning-based approach. In this work, we used regression to learn the properties of the solution space and have tested the trained algorithm on a number of examples to demonstrate the improvement gained. A general response model is constructed by learning from the annealing of benchmark circuits. This model is then used in the trained simulated annealing, which returns much better annealing quality than the untrained one for the same number of moves in the solution space. The annealing quality improvement was 15% ~ 43% for the set of examples used in training and 7% ~ 21% when the trained algorithm was applied to new examples. With the same amount of CPU time, the TSA could improve the annealing quality by up to 28% for some benchmark circuits we tested. In addition, the use of the response model successfully predicted the effect of the windowed sampling technique and derived the informally accepted advantages of windowing from the test set automatically.