Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
a r t i c l e in fo abstract
Article history: This paper investigates a hybrid evolutionary-based design system for automated sizing of analog
Received 29 July 2007 integrated circuits (ICs). A new algorithm, called competitive co-evolutionary differential evolution
Received in revised form (CODE), is proposed to design analog ICs with practical user-defined specifications. On the basis of the
28 January 2008
combination of HSPICE and MATLAB, the system links circuit performances, evaluated through electrical
Accepted 10 April 2008
simulation, to the optimization system in the MATLAB environment, once a circuit topology is selected.
The system has been tested by typical and hard-to-design cases, such as complex analog blocks with
Keywords: stringent design requirements. The results show that the design specifications are closely met, even in
Analog circuit synthesis
highly-constrained situations. Comparisons with available methods like genetic algorithms and
Analog circuit optimization
differential evolution, which use static penalty functions to handle design constraints, have also been
Differential evolution (DE)
Co-evolutionary differential evolution carried out, showing that the proposed algorithm offers important advantages in terms of optimization
(CODE) quality and robustness. Moreover, the algorithm is shown to be efficient.
Analog circuit sizing & 2008 Elsevier B.V. All rights reserved.
Corresponding author. 1
The maximization of a design objective can easily be transformed into a
E-mail address: wangy46@tsinghua.edu.cn (Y. Wang). minimization problem by just inverting its sign.
0167-9260/$ - see front matter & 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.vlsi.2008.04.003
ARTICLE IN PRESS
called competitive co-evolutionary differential evolution (CODE) algorithm. These techniques are available in some commercial
algorithm, is proposed to deal with this constrained opti- electrical simulators [26]. The drawbacks of deterministic opti-
mization problem. The algorithm has several novel features, mization algorithms are mainly in the following three aspects: (1)
which enable it to deal with large-scale and highly-con- they require a good starting point, (2) an unsatisfactory local
strained problems in an acceptable computation time with high minimum may be reached in many cases, and (3) they often
robustness. require continuity and differentiability of the objective function.
The evolutionary algorithm has been implemented in MATLAB Some researchers have tried to address these difficulties, such as
[25]. Evaluation of the objective function and constraints (values in Ref. [16], where a method to determine the initial point is
of f(x) and g(x) in Eq. (1)) is performed by using an electrical presented. Another approach is the application of geometric
simulator, HSPICE [26], for which an appropriate link with programming methods, which guarantee the convergence to a
MATLAB has been implemented. global minimum [11]. However, they require a special formulation
The structure of the paper is as follows. Section 2 reviews of design equations, which make them share many of the
related work and motivates the strategy of our optimization disadvantages of equation-based systems. Research efforts on
approach. The evolutionary algorithm used in this approach, stochastic search algorithms, especially evolutionary computation
differential evolution, and its implementation, are discussed in (EC) algorithms (genetic algorithms, differential evolution, genetic
Section 3. Section 4 formulates the competitive co-evolution programming, etc.) have begun to appear in literature in recent
approach to handle constraints in the differential evolution years [1,17–23]. Due to the ability and efficiency to find a
algorithm. Section 5 provides practical examples and benchmark satisfactory solution, genetic algorithms (GA) have been employed
tests to show the efficiency, effectiveness and advantages as optimization routines for analog circuits in both, industry
of the proposed approach. Comparisons with other common and academia. For problems with practical design specifications,
methods are also carried out. Finally, some concluding remarks most reported approaches use the penalty function method to
are given. handle constraints. Though these works have made a significant
progress, the optimization algorithms for analog circuit design
automation remain an active research area because of the
2. Related work following reasons:
Synthesis can be carried out by the following two different (1) GA is the most popular evolutionary algorithm, but its search
approaches: knowledge- and optimization-based. The basic ability and convergence rate have been criticized. It has also
idea of knowledge-based synthesis is to formulate design been proved that canonical GA cannot converge to the global
equations in such a way that given the performance character- optimum [27]. GA with elitism converges to the global
istics, the design parameters can be calculated [4–6]. In these optimum theoretically, but it is not always the case in
tools, the quality of the solutions in terms of both accuracy and practice. On the other hand, some other population-based
robustness is not acceptable since the very concept of knowledge- metaheuristics (PBMH) methods, such as swarm intelligence
based sizing forces the design equations to be simple. Other [28] and differential evolution [29] are attracting much
drawbacks are the large preparatory time/effort required to attention in the community of operations research because
develop design plans or equations, the difficulty in using them of their advantages over GAs. Their potentials in analog circuit
in a different technology, and the limitation to a fixed set of design automation still need to be exploited.
circuits. (2) The constraint handling problem is very important in analog
In optimization-based synthesis, the problem is translated into circuit design, especially for high performance circuits, which
function minimization problems that can be solved through are always highly constrained. Most reported synthesis
numerical methods. Essentially, they are based on the introduc- methods use penalty functions to handle constraints, and
tion of a performance evaluator within an iterative optimization few of them investigated solution algorithms for high
loop. The system is called equation-based when the performance performance design problems. In these methods, the con-
evaluator is based on equations capturing the behavior of a circuit strained optimization problem is transformed into an un-
topology [7–11]. However, creating the equations often consumes constrained one by minimizing the following function:
much more time than manually designing the circuit. In addition,
the simplifications required in the closed form analytical 0
X
n
f ðxÞ ¼ f ðxÞ þ wi hg i ðxÞi, (2)
equations cause low accuracy and incompleteness. On the i¼1
contrary, simulation-based methods do not rely on analytical
equations but on SPICE-like simulations to evaluate the circuit where the parameters wi are the penalty coefficients and
performances in the optimization process, which result in super- /gi(x)S returns the absolute value of gi(x) if it is negative, and
ior accuracy, generality, and ease of use [12–15]. Therefore, our zero otherwise. The results of the methods based on penalty
system is simulation-based. Through the link between HSPICE and functions are very sensitive to the penalty coefficients, and
MATLAB, the candidate parameters are transmitted from the may not meet the designer’s specifications in many cases.
optimization system to the simulation engine, and the circuit Small values of penalty coefficients drive the search outside
performances obtained by the electrical simulator are returned to the feasible region and often produce infeasible solutions,
the optimization system. The penalty to pay is a relatively long while imposing very severe penalties make it difficult to drive
computation time (compared to other methods), although, as the the population to the optimum solution. Usually, exact
experimental results in Section 5 demonstrate, it can be kept solutions are hard to find without tuning the penalty
within acceptable limits. coefficients for many times. Although several penalty strate-
Techniques for analog circuit optimization that appeared in gies have been developed [30,31], there has been no general
literature can be broadly classified into two main categories: rule for designing penalty coefficients till now.
deterministic optimization algorithms and stochastic search (3) Ability to handle large-scale design problems is still under
algorithms (evolutionary computation algorithms, simulated investigation. Most of the available methods can deal with
annealing, etc.). The traditional deterministic optimization meth- about 10–20 variables simultaneously, but analog circuits
ods mainly include steepest-descent algorithm and downhill with 30 or more unknown variables are common.
ARTICLE IN PRESS
3. Differential evolution and its implementation where indices r1 and r2 (r1, r2A{1, 2, y, NP}) are randomly chosen
and mutually different, and also different from the current index i.
The differential evolution (DE) algorithm and its implementa- The scaling factor F (FA(0, 1+)) controls the amplification of the
tion are introduced briefly in this section. The DE algorithm is differential variation (Xr1(t)Xr2(t)). Although F has not an upper
suitable for unconstrained problems, and it is also a basic limit, FA(0, 2] is commonly used. The population size NP must be
component in CODE. at least 4, so that the mutation operator can be applied. Vector
Differential evolution is a population-based evolutionary Xr0(t) is the base vector to be perturbed. There is a variety of
computation technique, which uses a simple differential operator mechanisms to select this base vector.
to create new candidate solutions, and a one-to-one competition After the mutation phase, the crossover operator is applied to
scheme to greedily select new candidates. Recently, DE has increase the diversity of the population. Thus, for each target
attracted much attention in various technical fields [32,33]. individual, a trial vector Ui(t) ¼ [ui,1(t), y, ui,d(t)] is generated as
The flow diagram of the DE algorithm is summarized in Fig. 1. The follows:
DE algorithm starts with the random initialization of a population of
(
individuals in the search space and works on the cooperative vi;j ðtÞ if ðrandðjÞpCRÞ or j ¼ randnðiÞ;
behaviors of the individuals in the population. At each generation, ui;j ðtÞ ¼ (6)
xi;j ðtÞ otherwise;
the mutation and crossover operators are applied to the individuals,
and a new population arises. Then, selection takes place, and the
where rand(j) is a random number uniformly distributed in the
corresponding individuals from both populations compete to build
range [0,1]. The index randn(i) is randomly chosen from the set {1,
the next generation. The algorithm tries to find the globally optimal
2, y, d}, and prevents the trial vector from being identical to the
solution by utilizing the distribution of solutions in the search space
target vector. The parameter CRA[0,1] is a constant called
and differences between pairs of solutions as search directions.
crossover parameter that controls the diversity of the population.
However, the searching behavior of each individual in the search
Following the crossover operation, the selection arises to decide
space is adjusted by dynamically altering the direction and step
whether the trial vector Ui(t) will be a member of the population of
length in which the search is performed.
the next generation t+1. For a minimization problem, Ui(t) is
The ith individual in the d-dimensional search space at
compared to the initial target individual Xi(t) by the following
generation t can be represented as
one-to-one based greedy selection criterion:
XðtÞ ¼ ½xi;1 ; xi;2 ; . . . ; xi;d ; i ¼ 1; 2; . . . ; NP, (3)
(
U i ðtÞ if f ðU i ðtÞÞof ðX i ðtÞÞ;
where NP denotes the size of the population. X i ðt þ 1Þ ¼ (7)
For each target individual i, according to the mutation operator, X i ðtÞ otherwise;
a mutant vector:
where Xi(t+1) is the new individual of the population of the next
V i ðt þ 1Þ ¼ ½vi;1 ðt þ 1Þ; . . . ; vi;d ðt þ 1Þ (4)
generation, and f(x) is the objective function.
is generated by adding the weighted difference between a pair of The procedure described above is considered as the standard
individuals, randomly selected from the population at generation version of DE. Several strategies of DE have been proposed,
t, to another individual, as described by the following equation: depending on the selection of the base vector to be perturbed, the
number and selection of the trial vectors and the type of crossover
V i ðtÞ ¼ X r0 ðtÞ þ FðX r1 ðtÞ X r2 ðtÞÞ, (5)
operators [32,33]. In our implementation, the base vector Xr0(t) is
selected to be the best member of the current population to share
its information among the individuals of the population and bias
Set Ranges DE Mutation solutions towards better vectors.
Special care has been taken for handling boundaries of the
parameters of the search space. Two classes of boundaries are
distinguished: hard and weak. Hard boundaries are those that
Initialization DE Crossover cannot be exceeded (e.g., a passive resistance cannot be negative
or the transistor gate length cannot be below the minimum value
allowed in the technological process), even if there are mathe-
DE Selection matically better solutions beyond those points. During the
Select Base execution of the DE algorithm, the overstepped individuals are
Vector set to the nearest bounds. Weak boundaries are those roughly
Update estimated to reasonably limit the search space. Although all
Parameters individuals in the initial population are selected within these
bounds, mutations during the evolution of the population may
yield individuals beyond those limits if better solutions are found.
No Unlike GA and particle swarm optimization (PSO), DE can deal
Reach Maximum with this problem. DE is also more effective as the accuracy of
Generations? local search is better than that of GA and PSO [29].
For some parameters it is also interesting to use logarithmic
scales to favor lower values of the parameters with large spans,
e.g., if a bias current spans over several decades, high values,
Yes hence high power consumption, will be favored if a linear scale is
used. Other parameters must be discretized, e.g., device sizes can
only change according to a given grid. In our implementation,
Output mutant vectors are allowed to vary continuously, to promote
diversity. However, the parameters are set to the nearest grid
Fig. 1. Flow diagram of the DE algorithm. value when evaluating the fitness of the individual.
ARTICLE IN PRESS
1X m
Though DE is very effective and efficient, it is not enough Fðx; r; hÞ ¼ f ðxÞ þ r ððg þ yi Þ2 y2i Þ. (10)
for the sizing of analog circuits. All EC algorithms them- 2 i¼1 i i
selves lack a mechanism to deal with the constraints of a
Here, x is the vector of decision variables, r is the vector of penalty
problem, which remains an open research area. However,
parameters, and mi ¼ riyi is the Lagrangian multiplier associated
there exist user-defined specifications for most analog circuit
with the ith constraint. This process is repeated until convergence.
design problems, and these constraints must be appropriately
The optimization objective is to find the saddle point (x0 , l0 ), such
handled.
that:
The use of penalty functions is the most common method, but
it is very sensitive to penalty coefficients and can hardly get Fðx0 ; l0 ÞpFðx0 ; l0 ÞpFðx0 ; l0 Þ. (11)
satisfactory results without proper penalty coefficients. Though
In the Lagrangian dual method, there is not such a saddle point
several penalty strategies [30,31] have been developed to improve
for non-convex problems. However, augmented Lagrangians
static penalty coefficients, there is no general rule to determine
addresses the problem by convexifying the objective function
proper penalty coefficients till now. Michalewicz describes
with quadratic penalty terms associated with the constraints [43].
the difficulties in each available penalty strategy in [34]. More-
For most practical problems, a saddle point always exists, and x0 is
over, Michalewicz and Schoenauer [35] concluded that the
the optimal solution of the optimization problem.
static penalty function method without any sophistication
is more robust, as one such sophisticated method may work
well on some problems but may not work well on another 4.2. Combination of co-evolutionary methodology and augmented
problem. Lagrangians
In this paper, constraints are handled by using the augmented
Lagrangian method, which transforms the constrained optimiza- The main problem of the augmented Lagrangian method is
tion problem into a problem amenable to the DE algorithm how to update the Lagrange multipliers so that it converges to the
described in Section 3. Penalty parameters in the augmented saddle point to avoid local optimization. A predetermined
Lagrangian formulation are automatically updated during execu- updating scheme may work well on some problems but may not
tion of the algorithm to reach the optimum point, hence avoiding work well on another problem. Co-evolution methodology, relying
the problems related to inappropriate settings of the penalty on the current evolution result of decision variables, solves this
parameters. Parameters are updated based on a co-evolution problem.
methodology. Competitive co-evolution method was inspired by observing
In recent years, co-evolution methodologies, including co- predator–prey relationship, where organisms adapt to each other
operative and competitive mode, have attracted much attention. in a dynamic environment. Groups are rewarded if they defeat
Methods based on cooperative mode mainly aim at unconstrained individuals that compete with them. Competitive co-evolution
optimization problem [36,37]. Cooperative co-evolution for con- strategy can be viewed as arm races of two groups [39,40]. To
strained optimization problems has been proposed in Ref. [38]. In arouse competition, two populations, whose values of fitness
this paper, the competitive co-evolution concept [39–41] and the functions are opposite to each other, must be generated.
modified DE algorithm based on the augmented Lagrangian In order to find the saddle point in augmented Lagrangians, the
method are combined to formulate a hybrid algorithm, CODE, problem can be formulated as
for constrained optimization problems in analog IC synthesis. We
will begin by introducing augmented Lagrangians, and then 1X m
minx maxm Fðx; r; hÞ ¼ f ðxÞ þ r ððg þ yi Þ2 y2i Þ, (12)
discuss the combination of augmented Lagrangians with compe- 2 i¼1 i i
titive co-evolution methodology.
where x is vector of design variables, and l is the vector of
Lagrangian multipliers. The purpose of the first population is to
4.1. Augmented Lagrangians
minimize F(x, r, h) with individuals in this population encoding
values of x and using a fixed l generated from the second
A constrained non-linear optimization problem can be ex- population. The evolution of this population tries to satisfy the
pressed as following property of the saddle point: F(x0 , l0 )pF(x, l0 ) in
minimize f ðxÞ Eq. (11). The second population aims to maximize F(x, r, h) with
its individuals encoding values of l and using a fixed x generated
g i ðxÞp0; i ¼ 1; . . . ; m
from the first population. The reason of this operation is F(x0 ,
subject to
l)pF(x0 , l0 ) in Eq. (11). This process establishes arm races of the
g j ðxÞ ¼ 0; j ¼ m þ 1; . . . ; n. (8) two populations. Once the first population achieves a solution x
These functions can be combined into a single transformation with a previous value of l, the second population gets a better l
function U, called the augmented Lagrangians [42]: based on x to defeat it. Then the first population generates a better
x to defeat the second population. At last, the saddle point in
1X m
1 X n
Eq. (11) can be reached, which is the optimal point of x. Values of r
Fðx; r; hÞ ¼ f ðxÞ þ r i ððg i þ yi Þ2 y2i Þ þ r ððg þ yj Þ2 y2j Þ.
2 i¼1 2 j¼mþ1 j j and h are initialized at the beginning of the optimization process,
and are updated at the end of each cycle of DE-based optimiza-
(9)
tion. The penalty coefficients should increase as ri+1 ¼ ri a after
In analog circuit design, equality constraints are limited to current each cycle, where r0 and a should be initialized first. The flow
and voltage relationships imposed by Kirchhoff’s laws: KCL and diagram of CODE is summarized in Fig. 2.
KVL. Kirchhoff’s laws are automatically included in the circuit In the first iteration of the DE-based optimization cycle, an
equations in electrical simulators, so only the inequality con- individual is randomly selected from the second population. Its
straints defined by design specifications have to be taken into purpose is that the first population needs a fixed Lagrangian
account in the optimization process. Therefore, the augmented multiplier from the second population in the first generation, but
ARTICLE IN PRESS
Initialize
population A
Initialize
population B
Start
DE operations
DE operations No
No
Reach generation B?
Reach generation A?
Yes Yes
Design
variable Multiplier
Yes
Best result
the second population does not start its evolution operations at a ¼ 1.5 for all the problems, except in the specific experiments
the same time. which demonstrate the low sensitivity of the results to these
It can be seen that the CODE algorithm fully inherits the parameter values. The design parameter search space is quite
advantages of the augmented Lagrangian method. The exact wide in all cases. Transistor lengths were allowed to vary between
solution can be achieved by augmented Lagrangian method, the minimum value allowed by the technological process to
whereas this is not true for the static penalty function method 10 mm. Transistor widths were changed between the minimum
[43]. In addition, the advantages of the differential evolution technology value to several hundreds of micrometers. Capacitor
algorithm, as described in Section 3, are also inherited. As any values and bias currents and voltages also had broad (yet
other stochastic optimization algorithm, it cannot be guaranteed reasonable) ranges.
that the global optimum solution is found for every problem, but The inputs to the system are a SPICE net list file containing the
the experimental results demonstrate that better solutions than structure, and user defined specifications. All the examples are
previous methods are obtained in all cases. run on a 2.4 GHz PC with 1 GB RAM, in the MATLAB environment.
Reported computation times include processing time in MATLAB,
the communication time between HSPICE and MATLAB, and the
5. Experimental results simulation time of HSPICE.
In this section, the developed algorithm will be applied to three 5.1. Example 1: Design of a two-stage amplifier
practical analog circuit sizing problems and four mathematical
benchmark problems. The three circuit sizing problems corre- The main purpose of this example is to test the capability of
spond to three amplifiers of increasing complexity. The purpose of CODE to handle constraints. A typical Miller-compensated two-
these examples is to test the ability of CODE to handle highly- stage amplifier, shown in Fig. 3, is chosen first to test the
constrained optimization problems, the ability to handle large algorithm. The technology used is a 0.25 mm CMOS process and
search spaces (large number of design parameters), its compar- the load capacitance CL is 30 pF. The design parameters are:
ison to other optimization algorithms and its low sensitivity to the transistor widths and lengths, compensation capacitor and bias
initial values of the optimization parameters. Finally, benchmark currents.
tests of the evolutionary computation field for constrained The first experiment tries to achieve the design objectives and
optimization are shown. constraints shown in Table 1. Appropriate matching constraints
In all the examples, the DE step size F is 0.8 and the crossover were established and appropriate operating region was ensured
probability CR is 0.8. The inner generations of population by imposing constraints like VDS/(VGSVTH)41 for NMOS transis-
encoding values of x is 80 for problems with less than 20 tors. The same experiment was tried with the standard genetic
variables, and 100 for other problems, and the generations of and differential evolution algorithms and using the static penalty
population encoding values of l is 80. The above parameters are function method to handle constraints (denoted GA+PF and
commonly used in DE based algorithms. We used r0 ¼ 1 and DE+PF, respectively). We tried to manually improve the penalty
ARTICLE IN PRESS
Table 2
Specifications and results of CODE, GA+PF and DE+PF
Table 3
Parameters of the two-stage amplifier
Table 4
Table 1
Experiments with different initial values of r0 and a
Specifications and results of CODE, GA+PF and DE+PF
Specifications Constraints r0 ¼ 2, a ¼ 2 r0 ¼ 1, a ¼ 3
Specifications Constraints CODE GA+PF DE+PF
DC gain (dB) X85 93.603 86.117
DC gain (dB) X70 76.48 72.601 80.659
GBW (MHz) X2.5 2.5072 3.3719
GBW (MHz) X2 2.068 4.523 2.0406
Phase margin (1) X55 59.708 57.402
Phase margin (1) X50 55.946 49.876 55.641
Output swing (V) X2 2.029 2.07
Output swing (V) X2 2.2017 2.1256 1.9182
CMRR (dB) X80 82.9 80.873
CMRR (dB) X70 90.01 70.99 70.018
PSRR (dB) X85 93.622 86.154
PSRR (dB) X70 76.571 74.658 80.802 pffiffiffiffiffiffi
pffiffiffiffiffiffi Noise ðnV= HzÞ p20 16.204 10.088
Input noise ðnV= HzÞ p60 53.653 50.302 57.004
Slow rate (V/ms) X1.8 1.9555 2.0359
Slow rate (V/ms) X1.5 1.5209 1.649 1.503
Power (mW) Minimize 1.0535 1.0217
Power (mW) Minimize 0.73118 2.215 1.1164
Total run time (s) 10957 11031
Total run time (s) 10097 10126 9865
Table 5 Table 7
Specifications and results of the CODE algorithm for the TCFC amplifier Results of the DE+PF algorithm (same specifications than Table 5)
Total run time (s) 12553 Total run time (s) 12486 12891 13762
Fig. 5. (a) Gain-boosted folded-cascode amplifier; (b) P amplifier and (c) N amplifier.
algorithm is not able to meet the performance specifications and experiments with different number of generations are shown in
several transistors are not in the correct operating region. Tables 10 and 11. Best, worst and average results were extracted
As stated above, an important advantage of the CODE from 15 runs of the algorithm. The results show that the proposed
algorithm, is that the competitive co-evolution adjusts the penalty algorithm for constrained optimization problems, CODE, is quite
coefficients to the appropriate values. As an illustration, Table 9 effective.
shows the evolution of the Lagrange multipliers for all constraints To show the advantages of CODE, the same benchmark
along several cycles of the co-evolutionary algorithm. It can be problems were tried with the DE+PF and GA+PF algorithms.
seen that in a few cycles, the multipliers converge to the right Different penalty coefficients were used in each of the 15 runs and
values to achieve a proper solution. the best result among them is shown in Table 12.2 The comparison
of computation times is shown in Table 13.
From the comparison, we can conclude that methods based on
5.4. Benchmark problems for constrained optimization
penalty functions are worse than the methods based on co-
evolution methods. In particular, unlike many previous works, in
In computer science, especially in evolutionary computation,
all cases, we use the same optimization parameters introduced
benchmark problems are of great importance to evaluate and
above, which have not undergone any specific calculation in view
compare different algorithms. Benchmark problems are tough,
and if an algorithm works well in benchmark problems, it is often
regarded as very effective for medium-sized optimization pro-
blems. Three highly-constrained benchmark problems [45], 2
This is the most favorable comparison for the GA+PF and DE+PF algorithms,
described in Appendix are tested first. The results for three since bad sets of penalty coefficients tend to decrease the mean value.
ARTICLE IN PRESS
Table 8 Table 11
Specifications and results of CODE, GA+PF and DE+PF Results of benchmark problems as a function of the number of generations in
CODE
Specifications Constraints CODE GA+PF DE+PF
Problem Generations
DC gain (dB) 480 112.17 79.995 69.953
GBW (MHz) 4250 256.88 253.93 250.34 500 1000 1500
Phase margin (1) 465 70.128 72.665 72.94
Gain margin o1 0.99952 0.90579 2.7761 G9
dm1a 41.2 19.162 19.269 14.627 Best 682.9565 682.8185 680.6900
dm2 41.2 5.9545 5.0727 0.54239 Average 684.4214 683.6447 682.3335
dm3a 41.2 15.856 26.476 26.554 Worst 689.3154 687.3919 686.9259
dm4a 41.2 4.8841 2.029e-6 4.6684e-8
dm5a 41.2 2.6609 2.3041 0.37412 G7
dm6a 41.2 8.6422 1.0149 2.6458 Best 25.2218 25.2109 24.8946
dm1bp 41.2 7.7727 4.9564 4.6678 Average 28.2832 26.1763 25.6317
dm3bp 41.2 2.8415 1.1078 0.23445 Worst 34.4057 28.9355 27.0023
dm5bp 41.2 18.277 4.8062 0.66502
dm6bp 41.2 5.581 2.0614 0.071441
dm8bp 41.2 12.483 0.00039371 0.96522
Table 12
dm10bp 41.2 4.263 8.2996 21.63
Results of benchmark problems
dm1bn 41.2 6.8212 1.6035 1.6045
dm3bn 41.2 6.2987 0.0082057 0.00012044
Method Problem
dm5bn 41.2 6.9101 0.0198 0.00031191
dm6bn 41.2 2.0146 0.00074711 0.035939
G1 G7 G9
dm8bn 41.2 4.5137 4.0393 4.4812
dm10bn 41.2 2.9358 0.00020514 7.5559e-5
GA+PF 13.7597 685.9112 30.9991
Power (mW) Minimize 4.004 15.508 0.87602
DE+PF 14.9583 683.0213 25.0172
Total run time (s) 5472 6124 5220 CODE 15.0000 680.6900 24.8946
Table 9 Table 13
Evolution of Lagrangian multipliers Computation time on benchmark problems
G1 Acknowledgments
Best 14.9501 14.9999 15.0000
Average 13.2311 14.8068 14.8821
Worst 11.4860 13.8913 13.9021
This work has been supported by National Natural Science
Foundation of China grant no. 60676012 and Special Funds for
ARTICLE IN PRESS
X
4 X
13 with
G1 ðxÞ ¼ 5x1 þ 5x2 þ 5x3 þ 5x4 5 x2i xi
i¼1 i¼5
10pxi p10; i ¼ 1; . . . ; 10.
2x2 þ 2x3 þ x11 þ x12 10p0, and the function value is: G7 ðxn Þ ¼ 24:3062091.
2x4 x5 þ x10 p0,
2x6 x7 þ x11 p0, Test problem 4
2x8 x9 þ x12 p0,
8x1 þ x10 p0, Minimize
8x2 þ x11 p0, G0 ðxÞ ¼ x1 x2
8x3 þ x12 p0, subject to:
with
x2 2x41 þ 8x31 8x21 2p0
0pxi p1; i ¼ 1; . . . ; 9 x2 4x41 þ 32x31 88x21 þ 96x1 36p0
0pxi p100; i ¼ 10; 11; 12
with
0px13 p1
0px1 p3; 0px2 p4.
The optimum solution is xn ¼ ð1; 1; 11; 1; 1; 1; 1; 1; 3; 3; 3; 1Þ and the
function value is G1(x) ¼ 15. The optimum solution is:
xn ¼ ð2:32952024; 3:17849288Þ
Test problem 2 and the function value is: G0 ¼ 5:508013271.
Minimize References
[9] P.C. Maulik, L.R. Carley, D.J. Allstot, Sizing of cell-level analog circuits using [41] H. Barbosa, A coevolutionary genetic algorithm for constrained optimization,
constrained optimization techniques, IEEE J. Solid-State Circuits 28 (3) (1993) IEEE Cong. Evolut. Comput. (1999) 1605–1611.
233–241. [42] R.T. Rockafellar, Convex Analysis, 1972.
[10] S. Wang, H. Gao, Z. Wang, The research of parameter optimization design [43] M.S. Bazaraa, H.D. Sherali, C.M. Shetty, Nonlinear Programming: Theory and
method for analog circuit, in: Proceedings of the Instrumentation and Algorithms, second ed., Wiley, New York, 1993.
Management Technology Conference, 2004, pp. 1515–1518. [44] X. Peng, W. Sansen, Transconductance with capacitances feedback compen-
[11] M. Hershenson, S. Boyd, T. Lee, Optimal design of a CMOS op-amp via sation for multistage amplifiers, IEEE J. Solid-State Circuits 40 (7) (2005)
geometric programming, IEEE Trans. Comput. Aid. Design of ICs 20 (1) (2001) 1514–1520.
1–21. [45] S. Koziel, Z. Michalewicz, Evolutionary algorithms, homomorphous
[12] W. Nye, D.C. Riley, A. Sangiovanni-Vincentelli, A.L. Tits, DELIGHT. SPICE: an mappings, and constrained parameter optimization, Evol. Comput. 7 (1)
optimization-based system for the design of integrated circuits, IEEE Trans. (1999) 19–44.
Comput. Aid. Design 7 (4) (1988) 501–519. [46] Z. Michalewicz, Evolutionary computation techniques for nonlinear program-
[13] R. Phelps, M.J. Krasnicki, R.A. Rutenbar, L.R. Carley, J.R. Hellums, Anaconda: ming problems, Int. Trans. Oper. Res. 1 (2) (1994) 223–240.
simulation-based synthesis of analog circuits via stochastic pattern search,
IEEE Trans. Comput. Aid. Design 19 (6) (2000) 703–717.
[14] F. Medeiro, R. Rodrı́guez-Macı́as, F.V. Fernández, R. Domı́nguez-Castro, J.L.
Huertas, A. Rodrı́guez-Vázquez, Global design of analog cells using statistical Bo Liu was born in Beijing, China, on September 23,
optimization techniques, Analog Integr. Circuits Signal Process. 6 (3) (1994) 1984. He is currently a senior undergraduate student in
179–195. Tsinghua University, Beijing, China. Since 2005, he has
[15] R. Castro-López, F.V. Fernández, O. Guerra, A. Rodrı́guez, Reuse Based been a Research Assistant at the Tsinghua National
Methodologies and Tools in the Design of Analog and Mixed-Signal Integrated Laboratory for Information Science and Technology,
Circuits, Springer, 2006. Beijing, China. Since 2007, he has been a Research
[16] G. Stehr, M. Pronath, F. Schenkel, H. Graeb, K. Antreich, Initial sizing of analog Assistant at the CAD Laboratory, Institute of Micro-
integrated circuits by centering within topology-given implicit specifications, electronics, Tsinghua University, Beijing, China. He is
in: Proceedings of the IEEE/ACM International Conference of Computer Aided also a member of Analog Design Automation Research
Design, 2003, pp. 241–246. Group of University of Sevilla, Sevilla, Spain. His
[17] S. Balkir, G. Dundar, G. Alpaydin, Evolution based synthesis of analog research interests include analog integrated circuit
integrated circuits and systems, in: Proceedings of the NASA/DoD Conference synthesis and evolutionary computation algorithms.
on Evolution Hardware, 2004.
[18] K. Takemura, T. Koide, H. Mattausch, T. Tsuji, Analog-circuit-component
Yan Wang received the B.S. and M.S. degrees in
optimization with genetic algorithm, in: Proceedings of the 47th IEEE
electrical engineering from Xi’an Jiaotong University,
International Midwest Symposium on Circuits and Systems, 2004,
Xi’an, China, in 1988 and 1991, respectively, and the
pp. 489–492.
Ph.D. degree in semiconductor device and physics from
[19] M. Barros, G. Neves, J. Guilherme, N. Horta, An evolutionary optimization
the Institute of Semiconductors, Chinese Academy of
kernel with adaptive parameters applied to analog circuit design, in:
Science, Beijing, China, in 1995. Since 1999, she has
Proceedings of the International Symposium on Signals, Circuits and Systems,
been a Professor with the Institute of Microelectronics,
2005, pp. 545–548.
Tsinghua University, Beijing, China. Her research
[20] C. Goh, Y. Li, GA automated design and synthesis of analog circuits with
focuses on semiconductor device modeling.
practical constraints, IEEE Cong. Evol. Comput. 1 (2001) 170–177.
[21] J. Koza, F. Bennett, D. Andre, M. Keane, F. Dunlap, Automated synthesis of
analog electrical circuits by means of genetic programming, IEEE Trans. Evol.
Comput. 1 (2) (1997) 109–128.
[22] G. Alpaydin, S. Balkir, G. Dundar, An evolutionary approach to automatic
synthesis of high performance analog integrated circuits, IEEE Trans. Evol. Zhiping Yu graduated from Tsinghua University, Beij-
Comput. 7 (3) (2003). ing, China, in 1967 with B.S. degree. He received his
[23] W. Kruiskamp, D. Leenaerts, Darwin: CMOS opamp synthesis by means of a M.S. and Ph. D degrees from Stanford University,
genetic algorithm, in: Proceedings of the ACM/IEEE Design Automation Stanford, CA, USA in 1980, and 1985, respectively. He
Conference, 1995, pp. 433–438. is presently the professor in the Institute of Microelec-
[24] Virtuoso NeoCircuit circuit sizing and optimization. Homepage: /http:// tronics, Tsinghua University, Beijing, China. From 1989
www.cadence.com/products/custom_ic/neocircuit/index.aspxS. to 2002, he has been a senior research scientist in the
[25] /http://www.mathworks.com/products/matlab/S. Dept. of Electrical Engineering in Stanford University,
[26] /http://www.synopsys.com/products/mixedsignal/hspice/hspice.htmlS. USA, while serving as the faculty member in Tsinghua.
[27] G. Rudolph, Convergence analysis of canonical genetic algorithms, IEEE Trans. He returned to Tsinghua full time since September
Neural Netw. 5 (1994) 96–101. 2002 and holds Pericom Microelectronics Professor-
[28] J. Kennedy, R. Eberhart, Y. Shi, Swarm Intelligence, Morgan Kaufmann ship (2002-2004) established by Pericom Semiconduc-
Publishers, San Francisco, 2001. tor Corp. in San Jose, USA. His research interests
[29] R. Storn, K. Price, Differential evolution—a simple and efficient heuristic for include device simulation for nano-scale MOSFETs, quantum transport in
global optimization over continuous spaces, J. Global Opt. 11 (1997) 341–359. nanoelectronic devices, compact circuit modeling of passive and active compo-
[30] J. Joines, C. Houck, On the use of nonstationary penalty functions to solve nents in RF CMOS, and numerical analysis techniques.
nonlinear constrained optimization problems with GAs, Proc. IEEE Int. Conf.
Evol. Comput. (1994) 579–584.
Leibo Liu received the B.S. and Ph.D. degrees in
[31] K. Deb, An efficient constraint handling method for genetic algorithm,
electronic engineering from Tsinghua University, Beij-
Comput. Methods Appl. Mech. Eng. 186 (2000) 311–338.
ing, China, in 1999 and 2004, respectively. He is
[32] K. Price, R. Storn. Differential evolution. Homepage: /http://www.ICSI.Ber-
currently an associate professor with the Institute of
keley.edu/storn/code.htmlS.
Microelectronics, Tsinghua University, Beijing, China.
[33] K. Price, R. Storn, J. Lampinen, Differential Evolution. A Practical Approach to
His research interests include reconfigurable processor
Global Optimization, Springer, 2005.
design and design methodologies, and system-level
[34] Z. Michalewicz, Genetic algorithms, numerical optimization, and constraints,
energy-aware design approaches for portable applica-
in: L. Eshelman (Ed.), Proceedings of the Sixth International Conference on
tions.
Genetic Algorithms, Morgan Kaufman, San Mateo, 1995, pp. 151–158.
[35] Z. Michalewicz, M. Schoenauer, Evolutionary algorithms for constrained
parameter optimization problems, Evol. Comput. 4 (1) (1996) 1–32.
[36] Y. Shi, H. Teng, Z. Li, Cooperative co-evolutionary differential evolution for
function optimization, ICNC (2) 2005, Lecture Notes in Computer Science,
2005, pp 1080–1088. Miao Li received the B.S. in electronic engineering from
[37] M.A. Potter, K.A. De Jong, A cooperative co-evolutionary approach to function Tsinghua University, PR China, in 2006. He was at the
optimization, in: Proceedings of the Third Parallel Problem Solving From Institute of Microelectronics of Tsinghua University
Nature, Jerusalem, Israel, 1994, pp. 249–257. from the September of 2006 to now. His research
[38] B. Liu, H. Ma, X. Zhang, B. Liu, A memetic cooperative co-evolutionary focuses on modeling of III–V compound semiconductor
differential evolution algorithm for constrained optimization problems, IEEE materials and devices.
Cong. Evol. Comput., 2007.
[39] M. Kirley, A co-evolutionary genetic algorithm for job shop scheduling
problems, in: Proceedings of 3rd International Conference on Knowledge-
Based Intelligent Information Engineering System, Australia, 1999, pp. 84–87.
[40] J. Paredis, Steps towards co-evolutionary classification neural networks,
Artificial Life 5, MIT Press, 1994.
ARTICLE IN PRESS
Zheng Wang is an undergraduate student in electronic F.V. Fernández got the Physics-Electronics degree from
engineering from Tsinghua University, Beijing, China. the University of Seville in 1988 and his Ph.D. degree in
1992. In 1993, he worked as a postdoctoral research
fellow at Katholieke Universiteit Leuven (Belgium).
Since 1995, he is an Associate Professor at the
Department of Electronics and Electromagnetism of
University of Sevilla. He is also a researcher at CSIC-
IMSE-CNM. His research interests lie in the design and
design methodologies of analog and mixed-signal
circuits. Dr. Fernández has authored or edited three
books and has co-authored more than 100 papers in
international journals and conferences. Dr. Fernández
is currently the Editor-in-Chief of Integration, the VLSI
Journal (Elsevier). He regularly serves at the Program Committee of several
international conferences. He has also participated as researcher or main
Jing Lu received the B.S. in electronic engineering from researcher in several National and European R&D projects.
Tsinghua University, China, in 2005. She was a
graduate student at the Institute of Microelectronics
of Tsinghua University from 2005 and will receive the
master degree in 2008. Her research focuses on
modeling of III–V compound semiconductor materials
and devices.