GA-SVM Optimization Kernel applied to Analog IC Design Automation Manuel Barros 1,2,3 Jorge Guilherme 1,2 Nuno Horta 1,3 1 Instituto de Telecomunicações, 2 Escola Superior de Tecnologia de Tomar, 3 Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal Phone: +351-218418093, Fax: +351-218418472, e-mail: fmbarros@ipt.pt; n.horta@ieee.org Abstract 1 — This paper presents a circuit/system level synthesis and optimization approach based on a learning scheme using Support Vectors Machines (SVMs) and evolutionary strategies applied to the design of analog and mixed-signal ICs. This approach combines the best qualities of these two techniques, a robust classification and regression method and a powerful global optimization. The SVM is used to dynamically model performance space and identify the feasible design space regions while at the same time the evolutionary techniques are looking for the global optimum. Finally, the proposed optimization-based approach is demonstrated for the design of some analog circuits using HSPICE as the evaluation engine. I. INTRODUCTION Nowadays, the microelectronics market trends present an ever-increasing level of integration with special emphasis in the production of complex mixed-signal systems-on-chip (SoC). A relevant emphasis is given to the process of analog integrated circuits (IC) design. The optimizations tools appear, naturally, as the key factor for the tremendous effort of finding the design parameters, which satisfy a complex, high-dimensional, multi-objective and multi-constrained problem. Thus, the previous approaches of analog synthesis tools such as [DELIGHT.SPICE [1], FRIDGE [2], AMGIE [3], and ANACONDA[4], APE[5], employed hybrid optimization algorithms. In general, optimization tools for analog circuits design employ a circuit analysis tool in the inner loop of the optimization to determine the circuit’s performance. This is pointed out as a very flexible solution when compared with other methodologies (equation-based, knowledge-based) because it accommodates to any type of circuit topology and accuracy, only, depending on simulator models. However, like other stochastic evolutionary algorithms they require a large number of fitness evaluations to get an adequate solution, resulting in an overall time- consuming process. A more practical approach to make optimization based methods more realistically is the use of modeling techniques 1 The work presented in this paper is partially supported by the PRODEP Program, under the action no.5.3. to approximate the fitness function, which is more time- efficient than a true fitness evaluation. The model presented through out this paper follows a supervised learning algorithm that is trained with historical simulation data gathered in each iteration of the EA. This method can be seen as a way of incorporating knowledge in the optimization process. The quality of the model has a large impact in the evolutionary process, so a good management of the evolutionary optimization process as well as the approximated model is fundamental. The impact of these approximated models in the evolutionary algorithms follows mainly two approaches. In one hand, the approximated model can be used to guide the genetic operators (selection, crossover and mutation) [7] or in the second case, as a real substitute of the true fitness evaluation [8], [12]. The most common approximation models based on evolutionary computation includes the Regression models, the Krigging models, Neural Networks [6] and Support Vector Machines (SVM) [11], [13]. The objective of this work is to find an efficient optimization approach to the automatic synthesis of analog and mixed-signal ICs using an on-line learning scheme based on Support Vectors Machines (SVMs) and in-house evolutionary optimization tool [GENOM] [9]. This approach tries to combine the best qualities of these proficient algorithms, a robust classification and regression method with a powerful global optimization. The paper is organized as follows: section II, gives an overview of the overall project and presents the design automation methodology. In section III, the increase in algorithm performance due to the introduction of feasibility models is demonstrated. Finally, the conclusions are drawn in section IV. II. A GA-SVM ON-LINE LEARNING ALGORITHM A. Project Overview The work presented in this paper is an improvement of the optimization tool called GENOM [9], which is based on an enhanced genetic algorithm (GA) implementation. In GENOM, the optimization kernel results from the implementation of a modified genetic algorithm with self