A Fuzzy Selection Based Constraint Handling Method for Multi-objective Optimization of Analog Cells Bo Liua., Francisco V. Fernándezb., Peng Gaoa. and Georges Gielena. a. ESAT-MICAS, Katholieke Universiteit Leuven, Leuven, Belgium e-mail: georges.gielen@esat.kuleuven.be b. IMSE, CSIC and University of Sevilla, Sevilla, Spain e-mail: pacov@imse.cnm.es Abstract—In this paper, a constraint handling technique for multi-objective optimization of analog cells is presented. Selection-based constraint handling and fuzzy membership functions are combined to construct a new constraint handling method, multi-objective fuzzy selection (MOFS). This is integrated with multi-objective sizing, which has the following two advantages: (1) enhances the effectiveness and efficiency of handling specifications distinctly; (2) specializes in solving multi-objective analog sizing problems mimicking the designer’s interactive design process, avoiding inflexibility of crisp constraint sizing methods. I. INTRODUCTION Driven by the increasing complexity of modern analog integrated circuits (ICs) and the ever-demanding time-tomarket pressures, analog sizing has evolved from a “pencil and paper” approach to a “simulation and optimization” approach [1]. Especially in recent years, Pareto-based multiobjective sizing [1]-[8] has received much attention. Analog sizing problems often require handling multiple noncommensurate and even conflicting goals. By multiobjective sizing, the trade-offs and sensitivity analysis between the different objectives can be explored. In multi-objective sizing, performances can either be set as objectives or constraints. An objective means that the designer is interested in its trade-offs with all other objectives in the whole performance space, which contains much information; while a constraint usually means that the performance needs to be “larger than” or “smaller than” a certain value, and no trade-off information is considered. Usually, only a part of the important performances are considered as objectives, and the others can be set as constraints. The reason is that the designer is often not interested in the trade-offs among all objectives. Moreover, there are typically other design decisions, such as restricting the transistors to operate in the saturation region, which are essentially defined as functional constraints. Consequently, constraint handling technology is very important in multi-objective sizing. Although a few multiobjective sizing approaches have appeared in literature [1][8], there are very few works that study the constraint This work is supported by a special bilateral agreement scholarship of the Katholieke Universiteit Leuven, Leuven, Belgium. This research was also supported by the TIC-2532 Project, funded by Consejería de Innovación, Ciencia y Empresa, Junta de Andalucía, Spain. handling technology. Hence, effective and efficient approaches need to be investigated. Apart from the crisp constraint handling, the designer’s intentions need to be integrated in. The designer is also interested in solutions in the neighborhood of the often arbitrarily chosen constraint boundaries. It is possible that the designer would like to know if the power vs. area tradeoff curve will be significantly improved when the phase margin is slightly lower than 60 (assuming the spec is phase margin > 60 ). If we define the trade-off between the objectives in the whole performance space as global tradeoff, the trade-off between the neighborhoods of the constraint boundaries can be defined as local trade-off. Currently, with crisp constraints, the only way to explore the local trade-offs is by repeatedly tuning the constraint boundaries. For example, the designer may first reduce the 60 phase spec to 58 and check the results. If satisfactory, he / she may further decrease the constraint boundary to 55 , which is the lowest requirement and compare the results; otherwise, 60 will be retained. However, more than one spec may be involved in the local trade-offs in a real design, and the combination of them may result in numerous tunings. To address the above problems, a new constraint handling method, named multi-objective fuzzy selection (MOFS), is proposed in this paper, whose purpose is to reflect the designer’s imprecise intentions, enabling them to correctly control the local trade-offs in a single run. The rest of the paper is organized as follows. Section II elaborates the MOFS method. Section III provides some practical experiments and comparisons to show the effectiveness of the proposed approach, and the concluding remarks are discussed in Section IV. II. A. MOFS METHOD Basic constraint handling mechanism There are some constraint handling methods for multiobjective analog sizing. One of them is that constraints are not considered in the optimization process, the Pareto- optimal front (POF) of competing objectives is first built, and infeasible points are discarded from the POF at the end. This method suffers from a considerable waste of computational effort and the density of points in the POF may be severely degraded. Another method is setting objectives of infeasible solutions to highly non-optimal values to prevent it from being selected in the successive generation [1]. This method finds many difficulties when all the candidates are infeasible in the first few generations (a common situation in analog sizing). The selection-based constraint handling method proposed by Deb [9] is very effective and robust. This method is used as the basic constraint handling mechanism by MOFS. Given two candidates in the population, there may be, at most, three situations: (1) Both solutions are feasible; (2) Both solutions are infeasible; (3) One solution is feasible, but the other is not. Accordingly, the selection rules are: (1) Given two feasible solutions, select the one with the better objective function value; (2) Given two infeasible solutions, select the solution with the smaller constraint violation; (3) If one solution is feasible and the other is not, select the feasible solution. n Constraint violation is calculated as | min( gi ( x), 0) | i 1 for {g1 ( x) 0, g 2 ( x) 0, , g n ( x) 0} . B. Fuzzy numbers and construction To avoid repeated tuning of the specification boundaries, fuzzy sets theory [10] is a possible solution. In this paper, fuzzy numbers (also called fuzzy sets) are integrated in the optimization process. A fuzzy number is a set of values, each having a membership ( x ) that reflects the degree of acceptability of the required value. ( x) 1 means that the value is fully acceptable, and ( x) 0 means that the value is not acceptable at all. The concept and construction of fuzzy numbers are detailed in [10]. By using fuzzy sets, the specifications of analog sizing are not “black and white” definitions but vague definitions, (e.g. 60dB is preferred, but 59dB is also acceptable), which is much closer to human intentions. A more important thing is this technology enable the designer to control the local tradeoffs, thus can generate a sizing result capturing the designer’s intention in one run. degradations when the POF can effectively be improved. If there is no improvement to the dominance or distribution, the relaxation of the constraints is of little sense, because the new solution is inferior to the solution that does not allow the tolerance. Thirdly, solutions that have many performances with ( x) 0 but close to ( x) 0 may appear, but such solutions can hardly be acceptable for the designer, as the degradations are too large. To address the third problem, besides the lower and upper bounds of the fuzzy number, a -cut [10] is introduced. A -cut set is a clear set from the fuzzy set, which can be expressed as A { x | A ( x ) } . Therefore, there are 4 levels to judge an obtained performance, as shown in Table 1. With this classification, the designer may set a positive number m to require more than m performances to reach the comparatively satisfactory value. Table 1. Fuzzy identifications of the performances Level of performances Judgment Fully satisfactory ( x) 1 Comparatively satisfactory Allowed but not satisfactory Not acceptable ( x) AND ( x) 1 ( x) 0 AND ( x) ( x) 0 The selection rule is finally defined as follows. Considering an analog multi-objective sizing problem with n constraints, out of which m are required to reach the comparatively satisfactory level, (1) Given two solutions with min( i ( x)) 0, i 1, , n , select the solution with the smaller constraint violation; (2) Given one solution with min( i ( x)) 0, i 1, ,n and one solution with min( i ( x)) 0, select the solution with min( i ( x)) 0 ; (3) Given two solutions with min( i ( x)) 0, i 1, ,n, compute the number of constraints with i ( x ) (noted m1 as , m2 . Compute mA min(m1 , m2 ) , mB max(m1 , m2 ) ) (3.1) if mA m , select the solution with the better dominance; (3.2) if mB < m, select the solution with the better n ( x) . i C. MOFS selection rules In order to integrate human flexibility in the selection rules, human intentions are first analyzed, which can be summarized as follows. Firstly, the relaxation of the constraints should be small, because the designer wishes to explore the local trade-offs around the specifications. Secondly, the designer only accepts some constraint i 1 (3.3) if mA m mB , select the solution with more constraints that reach the comparatively satisfactory level. D. Environmental selection In multi-objective sizing, not only good convergence is the optimization goal, but also the diversity of the points in the POF is important. Thus, environmental selection, that n ( ( x) ). Second, if the distance of an individual in the i i 1 obtained population by previous truncation method is less than a threshold value, it is replaced by the individuals with n large ( x) i . Suppose there are N individuals with i 1 distances below the threshold, then the N individuals with the largest aggregated membership function values will be selected to replace them. It can be seen that the algorithm call for a balance between distances between non-dominated solutions and higher acceptability of specifications. III. EXPERIMENTAL RESULTS AND COMPARISONS In this section, we combine MOFS with NSGA-II (but can be analogously combined with any other multi-objective optimization engine) and test them with a practical analog circuit sizing problem: the folded-cascode amplifier in Fig. 1. The load capacitance is 5 pF, and the technology used is a 0.18 m CMOS process with 1.8V power supply. The design variables are the transistor widths, the transistor lengths, and the bias current. All the examples are run on a Pentium IV PC at 3GHz with 2GB RAM and Linux operating system. The program is implemented in MATLAB and HSPICE is used as the performance evaluator. VDD Using fuzzy numbers to capture human intentions in single objective analog circuit sizing has appeared in literature, such as [12]. But most of them are weight-sumbased, whose drawbacks have been shown in [13]. Therefore, weighted-sum methods are not considered here, and MOFS and the selection-based method using crisp constraints are compared. The population size is 200, and 20 runs with independent different random numbers in NSGA-II optimization are performed for each experiment. A first experiment is performed with the high performance constraints in the second column of Table 2 and a second experiment with the low constraints in the fourth column. One typical solution is plotted. The power vs. area POF are shown in Fig. 2. The high and low specifications and the average values of the 200 individuals in the typical solution on each spec are shown in Table 2. Table 2. Experimental results with crisp constraints Performances constraints average constraints DC gain (dB) 60.05 60 55 50.03 50 40 GBW (MHz) Phase margin (º) 82.60 60 60 Output swing (V) 1.25 1.2 1 Slew rate 34.00 30 20 (V/ s ) dm1 17.64 1 1 dm3 1.81 1 1 1 1 dm5 3.52 1 1 dm7 2.42 dm9 1.56 1 1 dm11 7.56 1 1 Time (s) 1706 average 55.05 40.02 85.72 1.23 22.86 17.58 1.67 4.24 2.23 1.79 8.28 1672 MOFS is then applied. The fuzzy membership functions used are shown by solid lines in Fig. 3. Notice that this is an example of fuzzy numbers used in sizing, and that different designers can use different fuzzy numbers according to their own preferences. high specs low specs fuzzy specs m=3 fuzzy specs m=4 fuzzy specs m=5 20 18 16 sqrt(area)(um) considers both dominance and distance between individuals, is necessary [11]. The truncation method is the most commonly used approach for environmental selection in recent years [11]. A typical truncation method proceeds as follows: (1) Collect all the candidates that compete for the next generation and rank them on dominance. (2) If the number of non-dominated solutions exceeds the size of the population, then the non-dominated solutions are ranked by distance. The solutions with ranks larger than the population size are cut off. (3) If the number of non-dominated solutions is less than the population size, then fill in the available places with dominated solutions according to their quality on dominance and distribution. In the second case above (usually occurs in multiobjective sizing), a possible problem is that though some points are within the population size, their distances are also small. Thus, even when selected to the next generation, the points have not enough contribution to the distribution. On the other hand, there is potential to improve total membership function values under such condition. A revision has been made as follows. First, all the non-dominated solutions are ranked according to the aggregated membership function 14 12 10 8 0.5 0.55 0.6 0.65 0.7 power(mw) 0.75 0.8 0.85 Fig. 2. POF curves in the MOFS experiments m3A m4A mbpA Vcm ibb m11A m1A m10A m2A vip vout vin Vcm m9A m8A m7A m6A cl mbnA m5A VSS Fig. 1 Folded-cascode amplifier The specifications and the average values of the obtained solutions in one typical run are shown in Table 3. A diacritical tilde over a spec, n , means a fuzzy number around n, as defined in Fig. 3. The satisfactory levels are 58dB, 46MHz, 57 , 1.1V, 26 V / s for DC gain, GBW, phase margin, output swing and slew rate, respectively. Parameter m is set to 3, 4 and 5 for the above 5 fuzzy constraints on performances, which means that at least m of the 5 constraints should meet the satisfactory value. (a) (b) 0.5 0 0 0 20 40 GBW (e) 6070 0.5 0 0 20 40 60 8090 phase margin IV. 1 membership membership 1 0.5 0 0.5 0 20 40 60 80 DC gain (d) the 200 points in the POF) is 51.21dB, 46.04MHz, 85.16 , 1.13V, 30.21 V / s . Compared DC gain and GBW with the 4th column in Table 3, we can find the revision of fuzzy numbers directed the sizing. 1 membership 1 membership membership 1 (c) 0 0.5 1 1.51.8 output swing 0.5 0 0 20 40 50 slew rate Fig. 3. Fuzzy membership functions in the MOFS experiments Specs m DC gain (dB) GBW (MHz) Phase margin (º) Output swing (V) Slew rate (V/ s ) dm1 dm3 dm5 dm7 dm9 dm11 Time (s) opposite idea, as shown by the dotted line in Fig. 3, he / she can revise the fuzzy numbers easily to direct the multiobjective sizing. The comparatively satisfactory value is set to 53dB for DC gain and 46MHz for GBW. Parameter m is set to 4. This experiment result of a typical run (average of Table 3. Experimental results with fuzzy constraints constraints Average 1 Average 2 Average 3 3 4 5 60 57.08 58.13 58.12 50 42.03 42.01 46.03 60 85.43 82.45 83.51 1.2 1.24 1.18 1.23 30 28.30 26.25 30.05 1 1 1 1 1 1 17.51 1.63 5.38 2.53 1.37 8.72 1761 17.66 2.03 3.67 1.66 1.41 9.10 1805 17.65 1.94 3.88 2.16 1.43 8.34 1836 It can be seen from Fig. 2 that the advantage of MOFS is obvious. For example, when m=5, all the specs reach the comparatively satisfactory value, which are very near to the high requirements, but the POF is much better than that of the high requirements. When m=3, we can see that the POF is comparable to that of using the low requirements. But at least three out of the five specs reached the comparatively satisfactory level. The relaxations are reliable and reflect the designer’s intention by the fuzzy sets they constructed and the comparatively satisfactory value as well as the m value they set. To sum up, MOFS can automatically recognize the useful relaxation and control the degree of them according to the designer’s intentions, so the local trade-offs can be explored successfully in one sizing process, and no additional spec tuning is necessary. In the following, we show how the designer controls the local trade-off based on his / her intentions by revising the fuzzy sets. In the past experiments, we assumed that the GBW has a comparatively large acceptable region and DC gain has a smaller one. But if another designer has the CONCLUSIONS In this paper, the MOFS method has been proposed for constraint handling in multi-objective analog integrated circuit sizing problems. Thanks to the selection-based method and the fuzzy sets theory, MOFS can automatically recognize the useful relaxation and control the degree of them according to the designer’s intentions. Thus, the local trade-off can be achieved successfully in one sizing process, avoiding tedious manual tunings. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] B. D. Smedt and G. Gielen, “WATSON: Design space boundary exploration and model generation for analog and RF IC design”, IEEE Trans. Computer Aided-Design, pp. 213-224, 2003 F. D. Bernardinis and A. S. Vincentelli, “A methodology for systemlevel analog design space exploration”, Design, Automation, Test and Exhibition in Europe. pp. 676-677. 2004 S. Tiwary, P. Tiwary and R. Rutenbar, “Generation of yield-aware Pareto surface for hierarchical circuit design space exploration”, Design Automation Conference, pp. 31-36. 2006. T. Eeckelaert, T. McConaghy and G. Gielen, “Efficient multiobjective synthesis of analog circuits using hierarchical Pareto-optimal performance hypersurfaces”, Design, Automation, Test and Exhibition in Europe, pp. 1-6. 2005. G. Stehr, H. Graeb and K. Antreich, “Analog Performance Space Exploration by Normal-Boundary Intersection and by FourierMotzkin Elimination,” IEEE Trans. on Computer-Aided Design, Vol. 26, No. 10, pp. 1733-1748, Oct. 2007. D. Mueller, H. Graeb and U. Schlichtman, “Trade-off design of analog circuit using global attainment and “wave front” sequential programming”, Design, Automation, Test and Exhibition in Europe, pp. 1-6, 2007. R. A. Rutenbar, G. Gielen and J. Roychowdhury, “Hierarchical modeling, optimization, and synthesis of system-level analog and RF modeling”, Proceedings of the IEEE, pp. 640-669, 2007. R. Castro-López, E. Roca and F.V. Fernández, “Multimode Pareto fronts for reconfigurable analogue circuits,” IET Electronics Letters, Vol. 45, No. 2, Jan. 2009. K. Deb, S. Agrawal, A. Pratap and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II”, IEEE Transactions on Evolutionary Computation, pp. 182-197. 2002. T. J. Ross, “Fuzzy logic with engineering applications”, McGraw-Hill press, 1995. C. Coello Coello, G. Lamont and D. Veldhuizen, “Evolutionary algorithms for solving multi-objective problems (second edition)”, 2007, Springer B. Sahu and A.K Dutta, “Automatic synthesis of CMOS operational amplifiers: a fuzzy optimization approach”, 7th Asia and South Pacific Design Automation Conference, pp. 366-371, 2002. B. Liu, F. V. Fernández and G. Gielen, “Fuzzy Selection Based Differential Evolution Algorithm for Analog Cell Sizing Capturing Imprecise Human Intentions”, IEEE Congress on Evolutionary Computation, pp. 622-629, 2009.