Author:

Abstract:

This thesis describes computational intelligence techniques for the design automation of analog and high-frequency integrated circuits. Electronic design automation aims to obtain highly optimized design solutions (effectiveness) in an efficient manner (efficiency). Hence, effective and efficient global optimization is the goal of all the proposed computational intelligence algorithms presented in this thesis. Three important research areas are addressed in this thesis. The first research area is the global optimization of highly constrained problems. Its application is nominal analog circuit sizing with (very) high performance specifications. Considering the state of the art, the analog circuit sizing problems with very severe specifications cannot be solved either by manual design nor automatic synthesis. Two solution methods are proposed. For the case of hard (crisp) severe specifications that must be satisfied, a memetic evolutionary algorithm, MSOEA, is proposed. It integrates evolution operators from different evolutionary algorithms properly to enhance the constraint handling ability and the optimization quality. The experimental results show that MSOEA achieves the desired goals and has important advantages compared to the current analog sizing methods and some successful algorithms from the evolutionary computation (EC) field. For the case of investigating the circuit performances and some problems in practical analog circuit design, when crisp severe constraints are too rigid, two fuzzy analog sizing methods (FSBDE for single-objective analog circuit sizing and MOFSS for multi-objective analog circuit sizing) are proposed. The proposed fuzzy selection-based constraint handling mechanism integrates the human's flexibility and the EA's high optimization ability to obtain intelligent automatic analog circuit sizing methodologies. The useful relaxation of specifications can automatically be recognized by the method and the designer's intentions are correctly reflected by controlling the degree of relaxation. By using fuzzy analog circuit sizing, the designer avoids tedious manual tunings of the crisp specifications to iteratively improve the circuit performance, which is now achieved by a single fuzzy optimization.The second research area is the efficient global optimization in uncertain environments. The application is variation-aware analog circuit sizing. The starting points are that there are very few variation-aware analog circuit sizing methods which meet the requirements of being general, accurate and efficient enough in the electronic design automation field, and the lack of multi-objective efficient and scalable uncertain optimization methods in the computational intelligence field. Several solution methods are proposed. The ORDE algorithm for yield optimization and single-objective variation-aware analog circuit sizing uses a two-stage yield estimation process with ordinal optimization in the first stage, which determines the simulation effort for each candidate solution “intelligently”. A proposed random-scale differential evolution mutation operator maintains the diversity and performs combined global and local search, thus enhancing the convergence speed of the search engine. Experimental results show that ORDE can provide very accurate results with far less computational cost (an order of magnitude smaller) than the current Monte-Carlo-based method using infeasible pruning and Latin Hypercube Sampling techniques. The AYLeSS method for accurate yield estimation of analog circuits with high number of process variation variables combines Latin Supercube Sampling with a proposed partitioning rule for analog circuits. AYLeSS solves the limitation on the number of effective dimensions encountered by ranking-integrated quasi-Monte-Carlo, which to date was the best state-of-the-art method for analog circuit yield estimation. Experimental results show that AYLeSS is the current most efficient and general analog circuit yield estimation method. The MOOLP algorithm is proposed for multi-objective variation-aware analog sizing. A two-phase optimization mechanism is proposed, which provides the ground for appropriately including ordinal optimization and AYLeSS in the MOEA/D (the state-of-the-art multi-objective optimization method in the EC field) framework. Experimental results show that highly optimized Pareto fronts of analog circuits with a large number of process variation variables can be obtained in a practical time, which can hardly be achieved by the currently available methods. The third research area is the efficient global optimization of computationally expensive black-box problems. The application area is the automated design of mm-wave passive components and integrated circuits and complex antennas. There is to date no available method for mm-wave integrated circuit design automation and no efficient method for integrated passive components working at high frequencies (60GHz and above) and complex antenna global optimization. In the computational intelligence field, there is a lack of methods to efficiently and effectively address computationally expensive medium-scale black-box optimization problems. Solution methods are proposed in this thesis. The proposed MMLDE algorithm for small-scale computationally expensive optimization problems is the first efficient and general method for the synthesis of integrated passive components at high frequencies. MMLDE introduces Gaussian Process machine learning and hybrid methods into evolutionary computation, in order to construct a memetic surrogate model assisted evolutionary algorithm. MMLDE provides comparable results with standard EC methods, which are the current best framework in terms of solution quality, but achieves an order of magnitude speed enhancement. The GPMOOG method utilizes surrogate models in a different way and is the first practical (efficient, general and stable) multi-objective electromagnetic simulation-based design automation method. For medium-scale computationally expensive optimization problems, two methods are proposed. The EMLDE algorithm uses a decomposition approach, which reformulates the medium-scale problem into a hierarchical structure, and the ABGPDE algorithm is proposed to solve the low-dimensional but more complex expensive optimization problems. Applying EMLDE, the first mm-wave circuit design automation method is proposed, focusing on synthesizing linear amplifiers at high mm-wave frequencies. Experimental results show that EMLDE can provide solutions that are comparable to directly using standard EC methods with EM simulations as performance evaluation, which is the best framework in terms of solution quality, but uses nearly an order of magnitude less computational effort. The synthesis of a three-stage 100GHz amplifier is carried out in only 25 hours with a highly optimized design solution. At last, the GPEME algorithm is proposed to solve medium-scale expensive black-box optimization problems without any limitation on the problem structure (e.g. decomposable). The key ideas are (1) the development of a surrogate modeling-aware search mechanism in order to solve expensive optimization problems when very limited exact function evaluations are allowed (when a traditional high-quality surrogate model is difficult to build), (2) the use of dimension reduction techniques to solve the problems brought by the higher dimensionality. In contrast with the traditional surrogate model assisted evolutionary algorithm (SAEA) frameworks based on standard evolutionary algorithms (EAs), a new framework is developed, which achieves global optimization with the support of a corresponding simplified surrogate modeling. Sammon mapping is introduced to transform the decision variables from tens of dimensions to a few dimensions, in order to take advantage of Gaussian Process surrogate modeling in a low-dimensional space. Empirical studies on benchmark problems with 20, 30 and 50 dimensions show the high efficiency and effectiveness of GPEME. Compared to the state-of-the-art SAEAs, solutions with better or comparable quality can be achieved with 2 to 8 times less exact function evaluations. Applying GPEME, the first method for general mm-wave integrated circuit synthesis is presented and is demonstrated for the synthesis of a mm-wave power amplifier as example. The first general and efficient method for complex antenna design automation is also presented and demonstrated by applying GPEME. This research has started the research direction of simulation-based electromagnetic design automation and introduces new mathematical backgrounds as basis for future innovations in high-frequency circuit design automation and computational intelligence.