Hartley [Sz87] have shown that a Cauchy-Lorentz distribution (p(∆x) =T 2 +∆x)ismuchmoreefficient,2since the larger probabilities for large jumps allow the possibility of tunneling through barriers. Such aone-dimensional distribution can be produced by tan(π/2 ∗ (2r − 1)) (where r is a uniformly distributedr<strong>and</strong>om number between 0 <strong>and</strong> 1). This modification allows a much faster cooling schedule; it is sufficientto lower the temperature asT j = T 0 /j. (2)While the above methods address the question of step-size generation in terms of the generating distribution,the width of this distribution (overall magnitude of the step sizes) has not been addressed. Thespecification of this magnitude is generally pragmatic: the step-size (width of the distribution) is chosen sothat the process spends a “reasonable” amount of time sampling unfavored sections of the parameter space.This may be achieved by adjusting this width so that a certain fraction of the attempted events are notaccepted (typically 1/2). Likewise, the number of steps taken before lowering the temperature should belarge enough to ensure that the process samples the available section of parameter-space thoroughly. Thenumber which satisfies this is very case-dependent, so no general prescriptions can be given.The generalization of these methods to multivariate cases introduces several challenges. Firstly, it is notclear whether it is best to perform steps in the variables simultaneously or individually. If the steps aretaken simultaneously, there are problems in producing the required distributions, since there is no closedformmethod for producing a multi-dimensional Cauchy distribution. It is possible to use a product ofone-dimensional Cauchy distributions, however this requires a modification of the cooling schedule. Severalother schemes have been suggested to overcome these difficulties (e.g., Very Fast Simulated Reannealing [In89]<strong>and</strong> Threshold Annealing [Du90]). Many other modifications <strong>and</strong> new schemes have been proposed over thepast decade; the reader interested in further details is directed to e.g., [Mo90, Ki83, Sa91] <strong>and</strong> referencestherein.10.2 ImplementationofSAinGOSIAIn addition to the challenges involved in implementing Simulated Annealing (SA) to multi-dimensionalproblems, several issues inherent in the current Coulomb-<strong>excitation</strong> technique require special attention inthe adaptation of SA techniques. There are three issues which will be discussed in the following sections:(1) since the steps taken in the matrix elements may vary in magnitude, the recalculation of the correctionfactors must be periodically performed; (2) the different sensitivities of the cost function (χ 2 ) to the differentmatrix elements implies that the step sizes for different matrix elements must be independent; (3) due topossible dependencies of these sensitivities on the portion of the parameter space currently being sampled,these step-sizes must be periodically re-evaluated. The limits on each matrix element which are set in theME subsection (described in the GOSIA manual) are obeyed in the annealing process as well.In the present work, the method of Szu <strong>and</strong> Hartley has been followed for determining the step sizefor each matrix element. Steps in the individual matrix elements are generated simultaneously, in order tobetter account for correlations between matrix elements (i.e., large steps in a certain matrix element maybe strongly hindered, unless simultaneous steps in a correlated matrix element are taken) according to aCauchy-Lorentz distribution. This is achieved by choosing∆x i = σ i tan( π 2 (2r i − 1)), (3)where r i is a r<strong>and</strong>om number between 0 <strong>and</strong> 1 <strong>and</strong> the “width” array σ is an attempt at accounting for the verydifferent sensitivities of the matrix elements. The elements of this width array are initially set proportionalto 1/(dχ 2 /dx i ), <strong>and</strong> the widths are monitored <strong>and</strong> adjusted throughout the annealing procedure. The σ i areadjusted by recalculating the sensitivities periodically (at a user-specified interval). This re-calculation maybe performed in one of two methods: the “forward” calculation <strong>and</strong> the “forward-backward” calculation.The former method consists of evaluating dχ 2 /dx i using two values of x i - the current value, <strong>and</strong> a step inthe “forward” direction x i + ∆x i . The latter method involves the calculation of χ 2 at the points x i + ∆x i<strong>and</strong> x i − ∆x i . In both cases only the first derivative (slope) is evaluated. Since the value of χ 2 at the currentpoint is known, the “forward-backward” method involves twice as many χ 2 evaluations, <strong>and</strong> therefore takesroughly twice the execution time.T150
The average range covered by each matrix element is also tracked, <strong>and</strong> the maximum of the two widthestimates is used (in order to account for possible inaccuracies in the sensitivities, which are calculated in thefast approximation). An overall scaling factor is applied to the width array (which is otherwise normalizedto Σσ 2 i =1) <strong>and</strong> this scaling factor is adjusted to maintain a ratio of accepted steps to attempted steps of1/2. This ensures that the code spends a reasonable fraction of the time sampling unfavored regions.Since Coulomb-<strong>excitation</strong> analyses are all somewhat unique, there are several parameters which mustbe supplied by the user to determine the flow of the program. A basic underst<strong>and</strong>ing of the logic of theannealing routine is very helpful for determining the appropriate values of these parameters for a givenCoulomb <strong>excitation</strong> <strong>analysis</strong>. The structure of the routine is described here, <strong>and</strong> it is suggested that the usermonitor the progress of the procedure regularly.The core section of the annealing routine is a loop similar to that described above. That is, a sequence:Generate a step (in all matrix elements)calculate ∆χ 2 for this stepcalculate probability to accept step G(∆χ 2 ,T)accept or reject stepthat is executed a number of times controlled by the input quantities IACC_LIM <strong>and</strong> ITRY_LIM. If thenumber of attempted steps exceeds ITRY_LIM or the number of accepted steps exceeds IACC_LIM, theloop will terminate. Upon termination of this “inner” loop, the current statistics are printed to the logfile <strong>and</strong> the width array σ(x i ) is re-scaled by an overall factor in order to keep the ratio of attempted toaccepted steps as close to 1/2 as possible. Before re-scaling, the sensitivities for each matrix element maybe re-calculated on the basis of the input parameter ISENS.The “inner” loop described above is then re-entered <strong>and</strong> executed until one of the previously describedlimits is again reached. The “inner” loop described above is thereby enclosed within an “outer” loop asfollows:Execute “inner” loop until terminationprint statisticsif isens “outer” loops have been executed, re-calculate sensitivities<strong>and</strong> width arrayre-scale width arrayOnce a total of IACC_TOT_LIM steps have been accepted at the current temperature, the program willimmediately leave both inner <strong>and</strong> outer loops, lower the temperature, <strong>and</strong> begin the above process again. Theprogram will, however, execute this “outer” loop a maximum of OUTERLIM times <strong>and</strong> then halt executionwithout lowering the temperature. The OUTERLIM parameter is therefore a maximum value that shouldnot be reached under normal execution. This is assumed to be a number large enough that reaching thisvalue indicates a condition which requires human intervention (for example, if no steps are ever accepteddue to an improperly chosen limit).There is one other condition which will result in the program halting execution. This condition is definedby the “current” χ 2 value dropping below a minimum value CHIOK. The reason for this condition is simple:if the program happens upon a very good fit, there may be little reason to continue the procedure. The usermay check the resultant matrix-element values <strong>and</strong> decide whether the solution is reasonable or not beforecontinuing execution of the program.Since the SA approach lends itself to long (≈1 day) execution times, the implementation of this approachin the Coulomb <strong>excitation</strong> program GOSIA has been designed to be as fault-tolerant as possible. In theevent of a computer crash, the program can usually be re-started with little loss of time. During execution,several auxiliary files are therefore employed for the storage of information necessary for a quick re-start. Thecurrent set of matrix elements is written to disk periodically so that in the event of a crash, the calculationmay be resumed from the last recorded point. The current width array σ(x i ), temperature, <strong>and</strong> numbers ofsteps attempted <strong>and</strong> accepted are also written to disk so that the program can be somewhat system-faulttolerant. The current “best-fit” matrix elements are also tracked <strong>and</strong> written to disk, so that the user mayinspect the quality of the fits. The correction factors (for the “fast” calculation) are also written to diskperiodically. These can be corrupted, however, in the case of unexpected program termination (i.e., a crash).The correction factors should therefore always be re-calculated when re-starting GOSIA.151
- Page 1:
COULOMB EXCITATION DATA ANALYSIS CO
- Page 4 and 5:
10 MINIMIZATION BY SIMULATED ANNEAL
- Page 6 and 7:
1 INTRODUCTION1.1 Gosia suite of Co
- Page 8 and 9:
104 Ru, 110 Pd, 165 Ho, 166 Er, 186
- Page 13 and 14:
Figure 1: Coordinate system used to
- Page 15 and 16:
Cλ E =1.116547 · (13.889122) λ (
- Page 17 and 18:
Figure 2: The orbital integrals R 2
- Page 19 and 20:
2.2 Gamma Decay Following Electroma
- Page 21 and 22:
where :d 2 σ= σ R (θ p ) X R kχ
- Page 23 and 24:
Formula 2.49 is valid only for t mu
- Page 25 and 26:
à XK(α) =exp−iτ i (E γ )x i (
- Page 27 and 28:
important to have an accurate knowl
- Page 29 and 30:
3 APPROXIMATE EVALUATION OF EXCITAT
- Page 31 and 32:
with the reduced matrix element M c
- Page 33 and 34:
q (20)s (0 + → 2 + ) · M 1 ζ (2
- Page 35 and 36:
esults of minimization and error ru
- Page 37 and 38:
adjustment of the stepsize accordin
- Page 39 and 40:
approximation reliability improves
- Page 41 and 42:
Zd 2 σ(I → I f )Y (I → I f )=s
- Page 43 and 44:
4.5 MinimizationThe minimization, i
- Page 45 and 46:
X(CC k Yk c − Yk e ) 2 /σ 2 k =m
- Page 47 and 48:
However, estimation of the stepsize
- Page 49 and 50:
It can be shown that as long as the
- Page 51 and 52:
een exceeded; third, the user-given
- Page 53 and 54:
where f k stands for the functional
- Page 55 and 56:
x i + δx i Rx iexp ¡ − 1 2 χ2
- Page 57 and 58:
method used for the minimization, i
- Page 59 and 60:
OP,ERRO (ERRORS) (5.6):Activates th
- Page 61 and 62:
-----OP,SIXJ (SIX-j SYMBOL) (5.25):
- Page 63 and 64:
5.3 CONT (CONTROL)This suboption of
- Page 65 and 66:
I,I1 Ranges of matrix elements to b
- Page 67 and 68:
CODE DEFAULT OTHER CONSEQUENCES OF
- Page 69 and 70:
5.4 OP,CORR (CORRECT )This executio
- Page 71 and 72:
5.6 OP,ERRO (ERRORS)ThemoduleofGOSI
- Page 73 and 74:
5.7 OP,EXIT (EXIT)This option cause
- Page 75 and 76:
M AControls the number of magnetic
- Page 77 and 78:
5.10 OP,GDET (GE DETECTORS)This opt
- Page 79 and 80:
5.12 OP,INTG (INTEGRATE)This comman
- Page 81 and 82:
¡ dE¢dx1 ..¡ dEdx¢Stopping powe
- Page 83 and 84:
NI1, NI2 Number of subdivisions of
- Page 85 and 86:
5.13 LEVE (LEVELS)Mandatory subopti
- Page 87 and 88:
5.15 ME (OP,COUL)Mandatory suboptio
- Page 89 and 90:
Figure 10: Model system having 4 st
- Page 91 and 92:
ME =< INDEX2||E(M)λ||INDEX1 > The
- Page 93 and 94:
When entering matrix elements in th
- Page 95 and 96:
There are no restrictions concernin
- Page 97 and 98:
5.18 OP,POIN (POINT CALCULATION)Thi
- Page 99 and 100: 5.20 OP,RAW (RAW UNCORRECTED γ YIE
- Page 101 and 102: 5.21 OP,RE,A (RELEASE,A)This option
- Page 103 and 104: 5.25 OP,SIXJ (SIXJ SYMBOL)This stan
- Page 105 and 106: 5.27 OP,THEO (COLLECTIVE MODEL ME)C
- Page 107 and 108: 2,5,1,-2,23,5,1,-2,23,6,1,-2,2Matri
- Page 109 and 110: 5.29 OP,TROU (TROUBLE)This troubles
- Page 111 and 112: to that of the previous experiment,
- Page 113 and 114: To reduce the unnecessary input, on
- Page 115 and 116: OP,STAR or OP,POIN under OP,GOSI. N
- Page 117 and 118: 5.31 INPUT OF EXPERIMENTAL γ-RAY Y
- Page 119 and 120: 6 QUADRUPOLE ROTATION INVARIANTS -
- Page 121 and 122: *½P 5 (J) = s(E2 × E2) J ׯh¾
- Page 123 and 124: The expectation value of cos3δ can
- Page 125 and 126: where ē is an arbitratry vector. D
- Page 127 and 128: achieved using “mixed“ calculat
- Page 129 and 130: TAPE9 Contains the parameters neede
- Page 131 and 132: TAPE18 Input file, containing the i
- Page 133 and 134: 7.4.4 CALCULATION OF THE INTEGRATED
- Page 135 and 136: OP,EXITInput: TAPE4,TAPE7,TAPE9Outp
- Page 137 and 138: OP,ERRO0,MS,MEND,1,0,RMAXand the fi
- Page 139 and 140: 8 SIMULTANEOUS COULOMB EXCITATION:
- Page 141 and 142: 4, 3, 1kr88.corKr corrected yields
- Page 143 and 144: 0 Correction for in-flight decay ch
- Page 145 and 146: OP, ERRO Estimation of errors of fi
- Page 147 and 148: 9 COULOMB EXCITATION OF ISOMERIC ST
- Page 149: configurations with a probability e
- Page 153 and 154: SFX,NTOTI1(1),I2(1),RSIGN(1)I1(2),I
- Page 155 and 156: 11.2 LearningtoWriteGosiaInputsThe
- Page 157 and 158: (1.6 MeV)1.1 MeV0.75 MeV0.4 MeV0.08
- Page 159 and 160: Define the germaniumdetector geomet
- Page 161 and 162: Figure 15: Flow diagram for Gosia m
- Page 163 and 164: gosia < 2-make-correction-factors.i
- Page 165 and 166: Issue the commandgosia < 9-diag-err
- Page 167 and 168: At this point, it is suggested to c
- Page 169 and 170: calculation.) In this case, a copy
- Page 171 and 172: 4,-4, -3.705, 3,44,5, 4.626, 3.,7.5
- Page 173 and 174: 90145901459014590145901459014590145
- Page 175 and 176: .10.028921.10.026031.10.023431.10.0
- Page 177 and 178: 5,5,634,650,82.000,84.000634,638,64
- Page 179 and 180: ***********************************
- Page 181 and 182: *** CHISQ= 0.134003E+01 ***MATRIX E
- Page 183 and 184: CALCULATED AND EXPERIMENTAL YIELDS
- Page 185 and 186: 11.7 Annotated excerpt from a Coulo
- Page 187 and 188: 11.8 Accuracy and speed of calculat
- Page 189 and 190: 18,10.056,0.068,0.082,0.1,0.12,0.15
- Page 191 and 192: line 152 Eu 182 Tanumber (keV) (keV
- Page 193 and 194: 1.6 Normalization between data sets
- Page 195 and 196: 13 GOSIA 2007 RELEASE NOTESThese no
- Page 197 and 198: Matrix elements 500(April 1990, T.
- Page 199 and 200: 14 GOSIA Manual UpdatesDATE UPDATE2
- Page 201 and 202:
[KIB08]T.Kibédi,T.W.Burrows,M.B.Tr