chapter8.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

7 X
V
0 X
0.33 (within an intelligent agent, it was proposed that the interaction of a simple agent with the) 108 695.12 P
1.89 (environment could also give rise to intelligent behavior without associated knowledge) 108 677.12 P
-0.25 (being explicitly represented within the agent or the environment. Thus, the knowledge that) 108 659.12 P
0.32 (determines the intelligent behavior emer) 108 641.12 P
0.32 (ges out of the interaction of the simple agent and) 302.93 641.12 P
-0.03 (the environment, a computational theme that goes back at least to Simon \0501969\051 and is the) 108 623.12 P
0.33 (foundation of situated activity \050Greeno 1993; Clancey 1993\051. The techniques proposed in) 108 605.12 P
1.57 (this dissertation extend this work by demonstrating that interactions between a general) 108 587.12 P
-0.12 (problem solver and the task environment can also determine task-speci\336c modi\336cations of) 108 569.12 P
-0.02 (the problem solver) 108 551.12 P
-0.02 (\325) 198.34 551.12 P
-0.02 (s internal representations. I dubbed this method of AI emer) 201.67 551.12 P
-0.02 (gent intelli-) 484.38 551.12 P
0.17 (gence and I ar) 108 533.12 P
0.17 (gued that at the knowledge level, a knowledge-based AI agent and an agent) 175.89 533.12 P
1.8 (whose knowledge emer) 108 515.12 P
1.8 (ges are indistinguishable. Because emer) 224.62 515.12 P
1.8 (gent intelligent systems) 423.12 515.12 P
1.86 (appear similar to knowledge-based AI methods at the knowledge level but contain no) 108 497.12 P
0.04 (explicit task knowledge, the knowledge in these systems must emer) 108 479.12 P
0.04 (ge at the abstract level) 432.6 479.12 P
1.05 (from the interaction of the simple problem solver with the task environment. Following) 108 461.12 P
2.15 (the identi\336cation and characterization of emer) 108 443.12 P
2.15 (gent intelligence, the next four chapters) 339.38 443.12 P
(demonstrated this concept through experimentation.) 108 425.12 T
2.72 (Chapter 4 described an experiment where an evolutionary algorithm was able to) 126 395.12 P
0.97 (induce an appropriate neural network architecture for several tasks without task-speci\336c) 108 377.12 P
2.29 (knowledge. I ar) 108 359.12 P
2.29 (gued that the heuristics typically used by connectionist researchers to) 187.3 359.12 P
0.85 (induce an architecture are overly simplistic in their approach. GNARL, the evolutionary) 108 341.12 P
1.53 (algorithm used in these experiments, created appropriate architectures for several tasks) 108 323.12 P
0.16 (while inducing the parametric values simultaneously) 108 305.12 P
0.16 (. GNARL was shown to induce com-) 360.84 305.12 P
0.05 (plete networks for several language learning tasks in comparable time to a second method) 108 287.12 P
0.53 (that only induced weights for a designed architecture. In these experiments, the generali-) 108 269.12 P
0.14 (zation of the networks created by GNARL was consistently better) 108 251.12 P
0.14 (. In another experiment,) 424.68 251.12 P
0.94 (GNARL induced networks to perform a simple control task. These experiments demon-) 108 233.12 P
-0.23 (strated that two networks induced by GNARL might solve the same task in much the same) 108 215.12 P
(manner although their method of achieving the solution dif) 108 197.12 T
(fers signi\336cantly) 390.94 197.12 T
(.) 470.44 197.12 T
0.1 (The experiments of Chapter 4 illustrate emer) 126 167.12 P
0.1 (gent intelligence at a simple and straight-) 341.55 167.12 P
3.13 (forward level. Appropriate architectures for the networks are induced by repeatedly) 108 149.12 P
1.43 (manipulating them with representation-speci\336c operators and testing them in the actual) 108 131.12 P
-0.09 (task environment. When a network architecture appears that provides better problem solv-) 108 113.12 P
0.57 (ing accuracy than the average ability of the current population, it and its variants receive) 108 95.12 P
FMENDPAGE
%%EndPage: "146" 4
%%Page: "147" 4
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
(147) 522.01 712 T
108 90 540 703.12 R
7 X
V
0 X
1.36 (increasingly more of the population space and hence more subsequent testing. Because) 108 695.12 P
0.19 (there was no task-speci\336c knowledge available within the problem solver to determine an) 108 677.12 P
-0.04 (appropriate network architecture, this knowledge emer) 108 659.12 P
-0.04 (ged at the more abstract knowledge) 370.02 659.12 P
(level without explicit instantiation in the environment or the problem solver) 108 641.12 T
(.) 471.12 641.12 T
1.03 (Chapter 5 furthered the experiments of Chapter 4 by adding two representation-spe-) 126 611.12 P
1.66 (ci\336c operators that protected components of the FSAs evolving in the population. The) 108 593.12 P
0.08 (added operators made only syntactic manipulations of the solutions rather than the typical) 108 575.12 P
1.06 (semantic manipulations. The intention was that the evolutionary algorithm would deter-) 108 557.12 P
1.18 (mine which components of the FSA representation were crucial to solving the task and) 108 539.12 P
0.2 (protect them from manipulation by the other reproduction operators. Experimental results) 108 521.12 P
0.45 (showed that given the same experimental conditions, the evolutionary algorithm with the) 108 503.12 P
0.01 (additional operators identi\336ed solutions more quickly) 108 485.12 P
0.01 (. Rather than requiring an automated) 364.1 485.12 P
0.07 (analysis of each FSA to determine which components were crucial to solving the task and) 108 467.12 P
0.33 (thus should be protected from manipulation, the interaction of the evolutionary algorithm) 108 449.12 P
(and the task environment allowed the pertinent knowledge to emer) 108 431.12 T
(ge dynamically) 427.89 431.12 T
(.) 500.72 431.12 T
2.62 (In Chapter 6, automatic task decomposition was investigated, without an explicit) 126 401.12 P
1.06 (description of the properties of an appropriate decomposition inside the problem solver) 108 383.12 P
1.06 (.) 537 383.12 P
0.03 (Experiments in this chapter evolved modular LISP programs using GLiB, an evolutionary) 108 365.12 P
-0.03 (algorithm that evolves a population of LISP programs, to learn both the legal moves and a) 108 347.12 P
1.29 (competitive strategy for solving the T) 108 329.12 P
1.29 (owers of Hanoi problem and playing T) 294.52 329.12 P
1.29 (ic-T) 488.4 329.12 P
1.29 (ac-T) 507.55 329.12 P
1.29 (oe) 528.68 329.12 P
0.1 (against a rule-based expert. Again, two additional reproduction operators were introduced) 108 311.12 P
0.49 (into the basic evolutionary algorithm that probabilistically selected and de\336ned functions) 108 293.12 P
1.35 (from the programs in the population using only syntactic modi\336cations to the evolving) 108 275.12 P
0.93 (LISP programs. The appropriateness of each randomly de\336ned function was determined) 108 257.12 P
0.06 (by their subsequent proliferation through the population. Bene\336cial functions were propa-) 108 239.12 P
1.3 (gated to many population members while inappropriate modules died out with the pro-) 108 221.12 P
0.15 (grams that used them. The contents of the modules induced by the process appeared to be) 108 203.12 P
-0.16 (applicable to a broad number of situations within the limitations of representation and per-) 108 185.12 P
-0.09 (formance. An interesting side ef) 108 167.12 P
-0.09 (fect of these experiments is that induced functions de\336ned) 261.62 167.12 P
1.33 (task-speci\336c representational abstractions. Each new function, when appropriate for the) 108 149.12 P
0.49 (task, de\336ned a more abstract level of representation above the original language that was) 108 131.12 P
0.26 (tailored to the task. In these experiments, both the applicability of a particular module for) 108 113.12 P
FMENDPAGE
%%EndPage: "147" 5
%%Page: "148" 5
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
(148) 522.01 712 T
108 90 540 702.56 R
7 X
V
0 X
0.37 (solving the task and the abstraction it embodied emer) 108 694.56 P
0.37 (ged from the simple problem solver) 366.59 694.56 P
(interacting with the task environment.) 108 676.56 T
0.64 (The \336nal set of experiments, described in Chapter 7, took a closer look at the role of) 126 646.56 P
1.59 (the \336tness function in evolutionary algorithms. Standard objective \336tness functions are) 108 628.56 P
-0.2 (usually simple evaluation functions. However) 108 610.56 P
-0.2 (, the numerical value returned by these func-) 326.91 610.56 P
0.51 (tions is often arbitrary and is occasionally a liability when an accurate objective function) 108 592.56 P
0.5 (is not easily provided for the task. In addition, evaluation functions are often designed to) 108 574.56 P
1.38 (provide a speci\336c gradient to ensure the success of the learning process. In the experi-) 108 556.56 P
-0.21 (ments of this chapter) 108 538.56 P
-0.21 (, the \336tness function used in Chapter 6 was varied for the T) 207.17 538.56 P
-0.21 (ic-T) 488.4 538.56 P
-0.21 (ac-T) 507.55 538.56 P
-0.21 (oe) 528.68 538.56 P
1.07 (task to observe the dif) 108 520.56 P
1.07 (ference in quality in the discovered modular LISP programs. The) 217.97 520.56 P
0.82 (various experts used to evaluate the evolving programs included one that chose its posi-) 108 502.56 P
-0.22 (tion at random, one that was an optimal strategy for the task and one identical to the expert) 108 484.56 P
1.31 (used in the experiments of Chapter 6. A fourth \336tness function that used no expert but) 108 466.56 P
0.58 (instead used the population members to play against themselves in a competitive tourna-) 108 448.56 P
-0.06 (ment to determine relative rankings was also tested. This evaluation function merely iden-) 108 430.56 P
-0.21 (ti\336ed which of the two programs passed to it was the better player) 108 412.56 P
-0.21 (. The results showed that) 421.26 412.56 P
1.27 (the competitive tournament induced LISP programs that were more robust players than) 108 394.56 P
0.54 (each of the \336tness functions that used rule-based experts. I ar) 108 376.56 P
0.54 (gued that this resulted from) 405.93 376.56 P
1.08 (an ecology of strategies developing in the population which provided a pressure for the) 108 358.56 P
0.5 (evolving LISP programs to generalize their performance over a wider array of strategies.) 108 340.56 P
-0.06 (Equivalent generalization was not apparent in the programs evolved using the single, non-) 108 322.56 P
0.66 (deterministic rule-based strategies in the \336tness function. Thus, a progression toward the) 108 304.56 P
0.87 (implied goal of optimal T) 108 286.56 P
0.87 (ic-T) 234.31 286.56 P
0.87 (ac-T) 253.45 286.56 P
0.87 (oe performance emer) 274.58 286.56 P
0.87 (ged from the interaction between) 378 286.56 P
-0.2 (population members in the environment without an explicit gradient supplied in the \336tness) 108 268.56 P
(function.) 108 250.56 T
-0.2 (From the experimental results presented in this dissertation, some features of emer) 126 220.56 P
-0.2 (gent) 519.35 220.56 P
3.36 (intelligence as implemented by evolutionary algorithms can be identi\336ed. First, the) 108 202.56 P
1.41 (dynamics of evolutionary algorithms create a race condition that favors portions of the) 108 184.56 P
0.1 (search space that are dense with near) 108 166.56 P
0.1 (-solutions. For an of) 285.52 166.56 P
0.1 (fspring to be retained as a parent) 382.87 166.56 P
-0.11 (in the subsequent generation, it must perform the task about as well as an average member) 108 148.56 P
0.04 (of the population. Thus, an of) 108 130.56 P
0.04 (fspring\325) 250.23 130.56 P
0.04 (s performance must be comparable to its parent. The) 287.54 130.56 P
1.81 (of) 108 112.56 P
1.81 (fspring must also be within the search neighborhood of the parent where the search) 117.78 112.56 P
-0.26 (neighborhood is de\336ned relative to the reproduction operators. For subsequent of) 108 94.56 P
-0.26 (fspring to) 493.95 94.56 P
FMENDPAGE
%%EndPage: "148" 6
%%Page: "149" 6
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
(149) 522.01 712 T
108 90 540 702.56 R
7 X
V
0 X
0.05 (be equally viable, they must again be competitive with the average abilities of the popula-) 108 694.56 P
0.63 (tion. Thus, the reproduction operators must have again mapped to a comparable solution) 108 676.56 P
0.52 (within the same search neighborhood. Eventually) 108 658.56 P
0.52 (, a section of the search space begins to) 346.99 658.56 P
-0.28 (\336ll the population and the search is concentrated in that area. However) 108 640.56 P
-0.28 (, it is not necessarily) 442.85 640.56 P
-0.12 (the case that only a single area of the search space is represented in the population. Islands) 108 622.56 P
-0.14 (of distinct areas in the search space that are all dense with near) 108 604.56 P
-0.14 (-solutions can also develop.) 407.17 604.56 P
0.89 (The relative amount of space occupied by a particular area of the solution space will be) 108 586.56 P
-0.05 (proportional to the density of solutions in the neighborhood. The race then is for represen-) 108 568.56 P
0.74 (tational space in the population and thus for assigned credit and future manipulation. As) 108 550.56 P
-0.08 (demonstrated in Chapters 5 and 6, this feature of evolutionary algorithms can be exploited) 108 532.56 P
2.17 (to produce bene\336cial task-speci\336c syntactic or) 108 514.56 P
2.17 (ganizations that speed-up acquisition of) 340.79 514.56 P
0.46 (solutions and determine appropriate task-speci\336c modularizations and high-level abstrac-) 108 496.56 P
(tions without the use of explicit knowledge.) 108 478.56 T
0.19 (A second feature of emer) 126 448.56 P
0.19 (gent intelligence as implemented by evolutionary algorithms) 247.74 448.56 P
1.14 (is the opportunistic nature of the acquired structures. In each of the experiments of this) 108 430.56 P
0.93 (dissertation, the structures that evolved displayed little similarity with a human problem) 108 412.56 P
1.06 (solver) 108 394.56 P
1.06 (\325) 137.76 394.56 P
1.06 (s preconceptions of how to solve the task. The solutions often took advantage of) 141.09 394.56 P
1.64 (representational constructions that would be too subtle or too complex for a human to) 108 376.56 P
-0.04 (design. This was evident in the experiments with GNARL and the T) 108 358.56 P
-0.04 (ower of Hanoi experi-) 433.87 358.56 P
-0.19 (ment with GLiB. As the representation language became more general, the resulting struc-) 108 340.56 P
1.58 (tures became increasingly exploitative of non-standard representational techniques. For) 108 322.56 P
0.4 (instance, in the T) 108 304.56 P
0.4 (ic-T) 191.38 304.56 P
0.4 (ac-T) 210.52 304.56 P
0.4 (oe experiments with GLiB of Chapters 6 and 7, the) 231.65 304.56 P
2 F
0.4 (<test>) 483.77 304.56 P
0 F
0.4 ( por-) 516.62 304.56 P
1.81 (tion of the conditional in the representational language often contained numerous side) 108 286.56 P
1.08 (ef) 108 268.56 P
1.08 (fects. This invariably lead to representations that were non-normative for the problem) 117.1 268.56 P
1.34 (and extremely cumbersome to describe. As mentioned previously) 108 250.56 P
1.34 (, these representations) 431.06 250.56 P
0.72 (appear similar in spirit to the distributed representations of connectionist networks \050Hin-) 108 232.56 P
0.48 (ton, Rumelhart and McClelland 1986; Sejnowski and Rosenber) 108 214.56 P
0.48 (g 1987; Smolensky 1988\051) 414.96 214.56 P
-0.13 (which are also constructed using an opportunistic mechanism. This property is likely to be) 108 196.56 P
(present in all implementations of emer) 108 178.56 T
(gent intelligence.) 292.66 178.56 T
0.51 (Thirdly) 126 148.56 P
0.51 (, emer) 161.2 148.56 P
0.51 (gent intelligence relies on the task environment, implemented as the \336t-) 191.46 148.56 P
0.16 (ness function in evolutionary algorithms, to work as a \336lter for or) 108 130.56 P
0.16 (ganizations that are lim-) 423.28 130.56 P
-0.18 (ited relative to the rest of the population. The abstract features and opportunistic structures) 108 112.56 P
0.12 (within the population are not induced for their objective truth but instead for their relative) 108 94.56 P
FMENDPAGE
%%EndPage: "149" 7
%%Page: "150" 7
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
(150) 522.01 712 T
108 90 540 702.56 R
7 X
V
0 X
1.12 (usefulness in solving the task and their ability to bias future search toward increasingly) 108 694.56 P
-0.04 (better solutions. The more useful a structure or feature is in the context of the current pop-) 108 676.56 P
2.04 (ulation, the more variations of it that will be explored in subsequent iterations of the) 108 658.56 P
1.38 (search. This is distinct from the general optimization approach of some intelligent sys-) 108 640.56 P
-0.26 (tems. Rather than \336nding the optimal construction, these methods merely locate one that is) 108 622.56 P
1.21 (useful, i.e., suf) 108 604.56 P
1.21 (\336cient. If there are idiosyncracies to the structure within some tolerance,) 180.82 604.56 P
-0.13 (they are worked around. Often even these idiosyncracies turn out to be useful, as shown in) 108 586.56 P
0.39 (the T) 108 568.56 P
0.39 (owers of Hanoi experiment of Chapter 6. Hence these techniques are closer to satis-) 132.54 568.56 P
1.3 (\336cing systems than function optimizers. If optimality is a concern, it can be a criterion) 108 550.56 P
(inserted into the \336tness function for problems when satis\336cing is unacceptable.) 108 532.56 T
1.95 (In the future, studies which further illuminate the dif) 126 502.56 P
1.95 (ferences between knowledge-) 393.57 502.56 P
0.42 (based AI and emer) 108 484.56 P
0.42 (gent intelligence are needed. In this dissertation, I claimed that the ills) 199.29 484.56 P
2.09 (of knowledge-based AI originated from its commitment to explicit knowledge. I then) 108 466.56 P
0.41 (described and demonstrated a specialization of AI that avoided this commitment and still) 108 448.56 P
0.5 (appeared intelligent. This removes the need for explicit knowledge in the problem solver) 108 430.56 P
-0.12 (and hence the associated dif) 108 412.56 P
-0.12 (\336culties. However) 241.86 412.56 P
-0.12 (, I did not demonstrate that emer) 329.19 412.56 P
-0.12 (gent intelli-) 484.48 412.56 P
1.17 (gence can be used to solve the same problems as knowledge-based AI more quickly or) 108 394.56 P
0.43 (ef) 108 376.56 P
0.43 (\336ciently or that they truly circumvent the classic problems of knowledge-based AI sys-) 117.1 376.56 P
0.96 (tems. I have no doubt that there are problems for which emer) 108 358.56 P
0.96 (gent intelligent techniques) 411.48 358.56 P
0.2 (are bene\336cial and others for which it is not. The distinctions between these two classes of) 108 340.56 P
1.16 (problems should provide information about both the limitations of knowledge-based AI) 108 322.56 P
-0.1 (methodology and the practice of emer) 108 304.56 P
-0.1 (gent intelligence. Similarly) 289.48 304.56 P
-0.1 (, it is clear that not all the) 418.77 304.56 P
-0.24 (knowledge necessary to solve every problem will always be adequately accessible through) 108 286.56 P
0.69 (emer) 108 268.56 P
0.69 (gent mechanisms. The distinctions between what types of knowledge are accessible) 131.76 268.56 P
1.12 (and what types are not should also provide signi\336cant insight into these techniques and) 108 250.56 P
0.83 (possibly the nature of human intelligence. Other future work should include the bene\336ts) 108 232.56 P
0.02 (and limitations of inserting optimality constraints into the \336tness function and the identi\336-) 108 214.56 P
(cation of a theory of environmental feedback.) 108 196.56 T
1.23 (Knowledge-based AI techniques assume that the knowledge level is the appropriate) 126 166.56 P
-0.03 (level at which to model intelligent agents. This level of description is certainly convenient) 108 148.56 P
-0.19 (for testing a model of intelligence to ensure that it is consistent with itself and the observa-) 108 130.56 P
-0.29 (tions that gave rise to it. But the symbolic level favored by knowledge-based AI is not nec-) 108 112.56 P
5.17 (essarily the best level at which to formulate a computational intelligence. This) 108 94.56 P
FMENDPAGE
%%EndPage: "150" 8
%%Page: "151" 8
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
(151) 522.01 712 T
108 90 540 703.12 R
7 X
V
0 X
0.06 (straightforward approach to modeling intelligence presents numerous computational dif) 108 695.12 P
0.06 (\336-) 529.34 695.12 P
0.8 (culties when realistic problems are modeled due to the transduction of information from) 108 677.12 P
2.79 (the task environment to the problem solver) 108 659.12 P
2.79 (. Opportunistic methods that dynamically) 329.95 659.12 P
0.42 (acquire pertinent task information from the environment provide a more general method-) 108 641.12 P
-0.1 (ology for constructing intelligent systems ef) 108 623.12 P
-0.1 (fectively and circumvent our biases as design-) 318.8 623.12 P
1.96 (ers and observers. Emer) 108 605.12 P
1.96 (gent intelligence suggests a class of techniques that avoid the) 228.9 605.12 P
0.54 (modeling commitments of knowledge-based AI while still appearing \322as if\323 they contain) 108 587.12 P
2 (equivalent resources. Clarifying the distinctions between knowledge-based AI systems) 108 569.12 P
-0.03 (and emer) 108 551.12 P
-0.03 (gent intelligent systems that perform equivalently) 152.04 551.12 P
-0.03 (, such as those described in this) 389.29 551.12 P
0.99 (dissertation, will provide signi\336cant insight into the nature of human and computational) 108 533.12 P
0.19 (intelligence, the limitations and bene\336ts of computational modeling, and the biases of our) 108 515.12 P
(scienti\336c observations.) 108 497.12 T
FMENDPAGE
%%EndPage: "151" 9
%%Trailer
%%BoundingBox: 0 0 612 792
%%Pages: 8 1
%%DocumentFonts: Times-Roman
%%+ Times-Bold
%%+ Times-Italic