chapter4.ps

收集了遗传算法、进化计算、神经网络、模糊系统、人工生命、复杂适应系统等相关领域近期的参考论文和研究报告

493.81 592.66 475.65 567.15 2 L
N
7 X
90 450 6.81 5.44 511.97 567.15 G
0 X
90 450 6.81 5.44 511.97 567.15 A
90 450 6.81 5.44 475.65 567.15 G
90 450 6.81 5.44 475.65 567.15 A
4 X
90 450 6.81 5.44 511.97 541.64 G
0 X
90 450 6.81 5.44 511.97 541.64 A
4 X
90 450 6.81 5.44 475.65 541.64 G
0 X
90 450 6.81 5.44 475.65 541.64 A
4 X
90 450 6.81 5.44 493.81 592.66 G
0 X
90 450 6.81 5.44 493.81 592.66 A
491.08 611.28 454.76 636.79 2 L
N
454.76 611.28 491.08 636.79 2 L
N
491.08 636.79 491.08 611.28 2 L
N
454.76 636.79 454.76 611.28 2 L
N
472.92 662.3 491.08 636.79 2 L
N
472.92 662.3 454.76 636.79 2 L
N
90 450 6.81 5.44 491.08 636.79 G
90 450 6.81 5.44 491.08 636.79 A
90 450 6.81 5.44 454.76 636.79 G
90 450 6.81 5.44 454.76 636.79 A
4 X
90 450 6.81 5.44 491.08 611.28 G
0 X
90 450 6.81 5.44 491.08 611.28 A
4 X
90 450 6.81 5.44 454.76 611.28 G
0 X
90 450 6.81 5.44 454.76 611.28 A
4 X
90 450 6.81 5.44 472.92 662.3 G
0 X
90 450 6.81 5.44 472.92 662.3 A
90 450 71.73 103.01 454.76 576.69 A
(Evaluation) 389.86 592.9 T
(space) 402.57 583.91 T
0 F
(A) 129.62 574.61 T
(B) 185.13 550.04 T
114.72 384.63 533.46 446.81 R
7 X
V
4 12 Q
0 X
-0.45 (Figure 12:) 114.72 438.81 P
3 F
-0.45 (The competing conventions pr) 169.78 438.81 P
-0.45 (oblem \050Schaffer) 312.88 438.81 P
-0.45 (, Whitley and Eshelman 1992\051.) 387.37 438.81 P
0.8 (Bit strings A and B map to structurally and computationally equivalent networks that) 114.72 426.81 P
3.69 (assign the hidden units in differ) 114.72 414.81 P
3.69 (ent or) 284.32 414.81 P
3.69 (ders. Because the bit strings ar) 315.88 414.81 P
3.69 (e distinct,) 483.13 414.81 P
1.4 (cr) 114.72 402.81 P
1.4 (ossover is likely to pr) 124.26 402.81 P
1.4 (oduce an offspring that contains multiple copies of the same) 232.02 402.81 P
(hidden node, yielding a network with less computational ability than either par) 114.72 390.81 T
(ent.) 494.07 390.81 T
0 0 612 792 C
FMENDPAGE
%%EndPage: "72" 9
%%Page: "73" 9
612 792 0 FMBEGINPAGE
108 63 540 702 R
7 X
0 K
V
108 711 540 720 R
V
0 12 Q
0 X
(73) 528.01 712 T
108 90 540 702 R
7 X
V
0 X
0.28 (Finally) 126 694 P
0.28 (, when the architecture is variable, deception can occur when the parents dif) 159.21 694 P
0.28 (fer) 526.69 694 P
1.61 (topologically) 108 676 P
1.61 (. The types of distributed representations that might develop in a network) 170.52 676 P
0.82 (varies widely with the number of hidden units and the network\325) 108 658 P
0.82 (s connectivity) 420.71 658 P
0.82 (. Thus, the) 487.71 658 P
2.1 (distributed representations of topologically distinct networks have a greater chance of) 108 640 P
-0.28 (being incompatible parents. This further reduces the likelihood that crossover will produce) 108 622 P
(good of) 108 604 T
(fspring.) 144.76 604 T
0.25 (In short, for crossover to be a viable operator when evolving networks, the interpreta-) 126 574 P
1.51 (tion function must somehow compensate for all three types of deceptiveness described) 108 556 P
0.57 (above. This suggests that the complexity of an appropriate interpretation function should) 108 538 P
1.59 (more than rival the complexity of the original learning problem. Thus, the prospect of) 108 520 P
2 (evolving connectionist networks with crossover goes against current theory associated) 108 502 P
4.12 (with both genetic algorithms and connectionist networks. Better results should be) 108 484 P
0.1 (expected with reproduction operators that respect the uniqueness and potential incompati-) 108 466 P
(bility of the developing distributed representations.) 108 448 T
1 F
(4.3.2  Network Induction with Evolutionary Pr) 108 412 T
(ogramming) 346.99 412 T
0 F
1.66 (Unlike genetic algorithms, evolutionary programming de\336nes representation-depen-) 126 386 P
1.14 (dent operators that create of) 108 368 P
1.14 (fspring within a speci\336c behavioral locus of the parent \050see) 246.24 368 P
-0.08 (Figure 13\051. Because EP models evolution at the level of the behavior of unmixing species,) 108 350 P
2.19 (recombination of distinct population members is not considered. EP\325) 108 332 P
2.19 (s commitment to) 454.99 332 P
0.77 (mutation as the sole reproductive operator for searching over a space is preferable when) 108 314 P
-0.03 (there is no suf) 108 296 P
-0.03 (\336cient calculus to guide recombination by crossover) 175.33 296 P
-0.03 (, or when separating the) 425.19 296 P
(search and evaluation spaces does not af) 108 278 T
(ford an advantage.) 301.62 278 T
0.12 (Some previous EP systems have addressed the problem of evolving connectionist net-) 126 248 P
-0 (works. Fogel, Fogel and Porto \0501990\051 investigates training feedforward networks on some) 108 230 P
-0.14 (classic connectionist problems. McDonnell and W) 108 212 P
-0.14 (aagen \0501992\051 uses EP to evolve the con-) 348.15 212 P
1.52 (nectivity of feedforward networks with a constant number of hidden units by evolving) 108 194 P
0.47 (both a weight matrix and a connectivity matrix. Fogel \0501992a\051 and Fogel \0501993a\051 use EP) 108 176 P
1.9 (to induce three-layer fully-connected feedforward networks with a variable number of) 108 158 P
0.52 (hidden units that employ good strategies for playing T) 108 140 P
0.52 (ic-T) 372.27 140 P
0.52 (ac-T) 391.41 140 P
0.52 (oe. Saravanan \0501992\051 uses) 412.54 140 P
(an EP algorithm to evolve the weights of designed architectures.) 108 122 T
FMENDPAGE
%%EndPage: "73" 10
%%Page: "74" 10
612 792 0 FMBEGINPAGE
108 63 540 702 R
7 X
0 K
V
108 711 540 720 R
V
0 12 Q
0 X
(74) 528.01 712 T
108 90 540 702 R
7 X
V
0 X
0.15 (In most of the above studies, the mutation operator alters the parameters of network) 126 399.82 P
2 F
0.15 (h) 532.77 399.82 P
0 F
(by the function:) 108 381.82 T
0 10 Q
(\050EQ 10\051) 507.53 347.42 T
0 12 Q
0.45 (where) 108 307.69 P
3 F
0.45 (w) 140.75 307.69 P
0 F
0.45 ( is a weight,) 148.75 307.69 P
2 14 Q
0.53 (e) 211.52 307.69 P
0 12 Q
0.45 (\050) 217.67 307.69 P
2 F
0.45 (h) 221.66 307.69 P
0 F
0.45 (\051 is the error of the network on the task,) 228.89 307.69 P
2 F
0.45 (a) 426.28 307.69 P
0 F
0.45 ( is a user) 433.85 307.69 P
0.45 (-de\336ned pro-) 477.27 307.69 P
0.37 (portionality constant, and) 108 289.69 P
3 F
0.37 (N) 234.38 289.69 P
0 F
0.37 (\050) 242.38 289.69 P
2 F
0.37 (m) 246.38 289.69 P
0 F
0.37 (,) 253.28 289.69 P
2 F
0.37 (s) 259.65 289.69 P
0 F
0.37 (\051 is a gaussian variable with mean) 266.89 289.69 P
2 F
0.37 (m) 435.05 289.69 P
0 F
0.37 (and standard devia-) 445.33 289.69 P
(tion) 108 271.69 T
2 F
(s) 129.66 271.69 T
0 F
(.) 136.89 271.69 T
0.36 (The implementations of structural mutations in the studies that make structural mutations) 108 241.69 P
1.21 (dif) 108 223.69 P
1.21 (fer somewhat. McDonnell and W) 121.11 223.69 P
1.21 (aagen \0501992\051 randomly selects a set of weights and) 285.19 223.69 P
0.69 (alters their values with a probability based on the variance of the incident nodes\325 activa-) 108 205.69 P
1.02 (tion over the training set; connections from nodes with a high variance having less of a) 108 187.69 P
1.51 (chance of being altered. The structural mutation used in Fogel \0501992a; 1992b\051 adds or) 108 169.69 P
0.64 (deletes a single hidden unit with equal probability) 108 151.69 P
0.64 (. As in connectionist and some genetic) 351.25 151.69 P
0.39 (algorithm approaches to inducing network architectures, both of these EP studies employ) 108 133.69 P
(weak heuristics of modi\336cation that limit the accessibility of certain architectures.) 108 115.69 T
108 90 540 702 C
108 407.82 540 702 C
215.04 608.61 215.04 588.59 233.04 578.58 233.04 608.61 4 Y
5 X
0 K
V
1 H
2 Z
0 X
N
3 14 Q
(Structur) 132.04 675.95 T
(e) 177.39 675.95 T
(space) 141.04 662.61 T
7 X
270 360 15.75 21.45 208.29 582.87 G
0.5 H
2 X
270 360 15.75 21.45 208.29 582.87 A
7 X
180 270 25.45 21.45 208.99 582.87 G
183.74 570.68 183.54 582.87 190.17 572.64 186.6 572.83 4 Y
2 X
V
208 270 25.45 21.45 208.99 582.87 A
188.04 603.47 188.04 582.87 170.04 589.74 170.04 603.47 179.04 617.19 5 Y
12 X
V
0 X
N
2 X
90 450 58.5 55.77 206.04 591.93 A
261.06 620.7 255.21 622.07 259.32 626.45 3 L
0 Z
0 X
N
255.46 622.14 312.22 639.23 2 L
2 Z
N
(Locus of) 317.04 660.89 T
(mutation) 322.04 645.64 T
2 X
90 450 45 42.9 464.04 595.74 A
422.9 630.55 426.53 625.78 420.58 625.02 3 L
0 Z
0 X
N
394.22 639.23 426.31 625.87 2 L
2 Z
N
3 X
90 450 9 8.58 467.04 604.32 G
0 X
90 450 9 8.58 467.04 604.32 A
7 X
90 180 22.5 12.87 458.04 591.45 G
2 X
90 180 22.5 12.87 458.04 591.45 A
7 X
180 270 22.5 13.44 458.04 592.02 G
446.75 583.73 458.04 578.58 445.74 576.97 447.48 580.17 4 Y
2 X
V
180 242 22.5 13.44 458.04 592.02 A
3 X
90 450 9 4.29 467.04 578.58 G
0 X
90 450 9 4.29 467.04 578.58 A
(Mutation) 328.74 537.4 T
(operation) 334.74 524.05 T
429.59 582.05 435.53 582.88 433.28 577.32 3 L
0 Z
N
396.22 552.47 435.34 582.72 2 L
2 Z
N
224.1 561.82 219.54 565.72 225.19 567.72 3 L
0 Z
N
318.22 547.7 219.79 565.67 2 L
2 Z
N
124.04 510.2 524.54 695.85 R
N
120.99 449.14 527.59 488.88 R
7 X
V
4 12 Q
0 X
0.71 (Figure 13:) 120.99 480.88 P
3 F
0.71 (The evolutionary pr) 178.38 480.88 P
0.71 (ogramming appr) 274.64 480.88 P
0.71 (oach to modeling evolution. Unlike) 355.86 480.88 P
0.77 (genetic algorithm, evolutionary pr) 120.99 468.88 P
0.77 (ograms perform sear) 288.1 468.88 P
0.77 (ch in the space of networks.) 390.48 468.88 P
(Offspring cr) 120.99 456.88 T
(eated by mutation r) 179.51 456.88 T
(emain within a locus of similarity to their par) 272.68 456.88 T
(ents.) 490.8 456.88 T
108 90 540 702 C
0 0 612 792 C
202.88 339.69 412.65 359.82 C
3 12 Q
0 X
0 K
(w) 221.75 347.42 T
(w) 248.33 347.42 T
(N) 268.91 347.42 T
0 F
(0) 284.72 347.42 T
2 F
(a) 296.72 347.42 T
(e) 304.99 347.42 T
(h) 314.25 347.42 T
(\050) 310.26 347.42 T
(\051) 321.48 347.42 T
(,) 290.72 347.42 T
(\050) 279.62 347.42 T
(\051) 326.08 347.42 T
(+) 259.33 347.42 T
3 F
(w) 365.33 347.42 T
2 F
(h) 387.88 347.42 T
(\316) 376.33 347.42 T
(") 356.78 347.42 T
(=) 235.75 347.42 T
0 0 612 792 C
FMENDPAGE
%%EndPage: "74" 11
%%Page: "75" 11
612 792 0 FMBEGINPAGE
108 63 540 702 R
7 X
0 K
V
108 711 540 720 R
V
0 12 Q
0 X
(75) 528.01 712 T
108 90 540 702 R
7 X
V
0 X
-0.13 (Evolutionary programming provides distinct advantages over genetic algorithms when) 126 694 P
0.14 (evolving networks. First, EP manipulates the network representation directly) 108 676 P
0.14 (, thus obviat-) 476.76 676 P
0.69 (ing the need for a dual representation and the associated interpretation function. Second,) 108 658 P
1.14 (by avoiding crossover between networks in creating of) 108 640 P
1.14 (fspring, the individuality of each) 378.55 640 P
2.37 (network\325) 108 622 P
2.37 (s distributed representation is respected. For these reasons, evolutionary pro-) 150.64 622 P
0.97 (gramming provides a more appropriate framework for simultaneous structural and para-) 108 604 P
2.69 (metric learning in recurrent networks. But rather than supply simplistic heuristics to) 108 586 P
0.11 (signify when to manipulate the architecture, the interaction with the task environment can) 108 568 P
1.45 (fully specify an appropriate architecture. The GNARL algorithm, presented in the next) 108 550 P
(section and investigated in the remainder of this chapter) 108 532 T
(, describes one such approach.) 375.99 532 T
1 F
(4.4  The GNARL Algorithm) 108 496 T
0 F
1.12 (GNARL, which stands for) 126 472 P
3 F
1.12 ( GeNeralized Acquisition of Recurr) 255.95 472 P
1.12 (ent Links) 429.9 472 P
0 F
1.12 (, is an evolu-) 474.67 472 P
1.12 (tionary algorithm that non-monotonically constructs recurrent connectionist networks to) 108 454 P
-0.04 (solve a given task. The name GNARL re\337ects the types of networks that arise from a gen-) 108 436 P
0.48 (eralized network induction algorithm performing both structural and parametric learning.) 108 418 P
-0.03 (Instead of having uniform or symmetric topologies, the resulting networks have \322gnarled\323) 108 400 P
0.85 (interconnections of hidden units which may accurately re\337ect the constraints inherent in) 108 382 P
(the task.) 108 364 T
0.56 (The general architecture of a GNARL network is straightforward. The input and out-) 126 334 P
0.53 (put nodes are considered to be provided by the task and are immutable by the algorithm;) 108 316 P
0.07 (thus each network for a given task always has) 108 298 P
3 F
0.07 (m) 331.44 298 P
3 10 Q
0.06 (in) 340.1 295 P
0 12 Q
0.07 (input nodes and) 350.43 298 P
3 F
0.07 (m) 429.59 298 P
3 10 Q
0.06 (out) 438.25 295 P
0 12 Q
0.07 (output nodes. The) 453.58 298 P
0.54 (number of hidden nodes varies from 0 to a user) 108 280 P
0.54 (-supplied maximum) 339.44 280 P
3 F
0.54 (h) 439.8 280 P
3 10 Q
0.45 (max) 445.79 277 P
3 12 Q
0.54 (.) 462.44 280 P
0 F
0.54 (Bias nodes are) 468.98 280 P
0.59 (optional; if provided in an experiment, they are implemented as an additional input node) 108 262 P
2.2 (with constant value one. All non-input nodes employ the standard sigmoid activation) 108 244 P
(function. Links use real-valued weights, and must obey three restrictions:) 108 226 T
3 F
(R) 162 196 T
3 10 Q
(1) 169.33 193 T
0 12 Q
(: There can be no links) 174.32 196 T
3 F
(to) 286.91 196 T
0 F
( an input node.) 296.24 196 T
3 F
(R) 162 172 T
3 10 Q
(2) 169.33 169 T
0 12 Q
(: There can be no links) 174.32 172 T
3 F
(fr) 286.91 172 T
(om) 294.46 172 T
0 F
( an output node.) 309.12 172 T
3 F
(R) 162 148 T
3 10 Q
(3) 169.33 145 T
0 12 Q
(: Given two nodes) 174.32 148 T
3 F
(x) 264.93 148 T
0 F
( and) 270.26 148 T
3 F
( y) 290.58 148 T
0 F
(, there is at most one link from) 298.12 148 T
3 F
(x) 448.36 148 T
0 F
( to) 453.69 148 T
3 F
(y) 469.02 148 T
0 F
(.) 473.56 148 T
-0.39 (Thus GNARL networks may have no connections, sparse connections, or full connectivity) 108 118 P
-0.39 (.) 537 118 P
1.48 (Consequently) 108 94 P
1.48 (, GNARL) 173.18 94 P
1.48 (\325) 220.86 94 P
1.48 (s search space is all networks,) 224.2 94