chapter1.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

0.17 ( This principle dictates that if the agent contains knowledge that is) 219.97 172 P
-0 (applicable in the current context and will satisfy one of its goals, then a rational agent will) 108 154 P
(apply the knowledge to achieve the goal.) 108 136 T
FMENDPAGE
%%EndPage: "2" 3
%%Page: "3" 3
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
(3) 534 712 T
108 90 540 702 R
7 X
V
0 X
1.28 (The simplest manner in which to model observed intelligent behavior is to take the) 126 694 P
-0.07 (observations at the knowledge level at face value and implement them directly at the sym-) 108 676 P
-0.29 (bol level. All that is needed is a representation for knowledge and goals and an algorithmic) 108 658 P
-0.05 (version of the principle of rationality) 108 640 P
-0.05 (. Knowledge and goals are extracted through analysis) 283.86 640 P
0.46 (of the task environment. After the knowledge and goals are extracted, weak methods can) 108 622 P
-0.08 (serve as an appropriate principle of rationality) 108 604 P
-0.08 (, assuming that control knowledge appropri-) 327.89 604 P
0.17 (ate for the task is added for ef) 108 586 P
0.17 (\336ciency) 251.82 586 P
0.17 (. Newell echoes this modeling approach by suggest-) 289.02 586 P
1.42 (ing that the knowledge level can be reduced to the symbol level. In this reduction, the) 108 568 P
0.97 (knowledge and goals at the knowledge level correspond simply to data structures in the) 108 550 P
(program:) 108 532 T
(When [AI researchers] say) 162 502 T
(, as we often do in explaining an action) 289.42 502 T
(of a program, that the \322program knows K\323 ... we mean that there is) 162 484 T
-0.17 (some structure in the program that we view as holding K... \050Newell) 162 466 P
(1981, p. 15\051) 162 448 T
(The discovery) 162 423 T
(, development and elaboration of [the symbol level]) 229.5 423 T
(to predicting the behavior of an intelligent agent has been what AI) 162 405 T
(has been all about... \050Newell 1981, p. 12\051) 162 387 T
1.26 (I label such models of intelligence, where there is a direct correspondence between the) 108 357 P
0.56 (knowledge described at the knowledge level and the content of the representations at the) 108 339 P
-0.25 (symbol level, as) 108 321 P
2 F
-0.25 (literal knowledge level models) 187.86 321 P
0 F
-0.25 ( of intelligent behavior) 333.33 321 P
-0.25 (. Such models repre-) 441.51 321 P
(sent the majority of knowledge-based AI systems.) 108 303 T
-0 (It is important to note that what is observed at the knowledge level may not accurately) 126 273 P
-0.23 (re\337ect the process that gave rise to the observations. In other words, there may be a dispar-) 108 255 P
0.74 (ity between the) 108 237 P
2 F
0.74 (appar) 186.5 237 P
0.74 (ent) 214.71 237 P
0 F
0.74 ( knowledge of the agent as seen at the knowledge level and the) 229.36 237 P
2 F
2.66 (actual) 108 219 P
0 F
2.66 (knowledge contained in the implementation of the agent at the symbol level.) 143.64 219 P
2.43 (Knowledge-based AI has long confused the implementation of intelligence within an) 108 201 P
0.17 (agent with its knowledge level description, choosing to equate the two to simplify model-) 108 183 P
2.44 (ing \050Chandrasekaran, Goel and Allemang 1989\051. Situated activists and connectionists) 108 165 P
0.06 (have noticed a similar confusion in the terms like) 108 147 P
2 F
0.06 (plan) 348.05 147 P
0 F
0.06 (,) 369.37 147 P
2 F
0.06 (symbol) 375.43 147 P
0 F
0.06 ( and) 409.4 147 P
2 F
0.06 (r) 432.84 147 P
0.06 (epr) 437.06 147 P
0.06 (esentation) 452.6 147 P
0 F
0.06 (as used) 504.97 147 P
2.17 (in knowledge-based AI systems \050W) 108 129 P
2.17 (inograd and Flores 1986; Hinton, McClelland and) 287.43 129 P
2.66 (Rumelhart 1986; Smolensky 1988; Greeno and Moore 1993; Clancey 1993\051. Brooks) 108 111 P
FMENDPAGE
%%EndPage: "3" 4
%%Page: "4" 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
(4) 534 712 T
108 90 540 702 R
7 X
V
0 X
1.6 (\0501986, 1991\051 makes similar points in his approach to modeling intelligent behavior by) 108 694 P
1.18 (constructing robots devoid of explicit representations. All of these confusions and criti-) 108 676 P
-0.26 (cisms, in my opinion, can be traced to knowledge-based AI\325) 108 658 P
-0.26 (s dependence on literal knowl-) 393.81 658 P
(edge level models of intelligence.) 108 640 T
1 F
(1.2  The Pr) 108 604 T
(oblems of Explicit Knowledge) 164.42 604 T
0 F
1.1 (In a literal knowledge level model, the knowledge to solve the problem is) 126 580 P
2 F
1.1 (explicitly) 496.03 580 P
0 F
-0 (represented within the problem solving agent in a knowledge-base using some representa-) 108 562 P
2.03 (tion. This dissertation de\336nes) 108 544 P
2 F
2.03 (explicit knowledge) 260.38 544 P
0 F
2.03 ( to be any manner of domain theory) 352.02 544 P
2.03 (,) 537 544 P
0.96 (domain knowledge, world model, heuristics, knowledge-base, or other form of informa-) 108 526 P
1.36 (tion,) 108 508 P
2 F
1.36 (r) 134.02 508 P
1.36 (egar) 138.24 508 P
1.36 (dless of r) 159.78 508 P
1.36 (epr) 206.04 508 P
1.36 (esentation) 221.58 508 P
0 F
1.36 (, that is speci\336c to the task being solved and included) 270.89 508 P
2 F
-0.19 (explicitly within) 108 490 P
0 F
-0.19 ( the problem solver) 184.78 490 P
-0.19 (. Explicit knowledge is the basis of strong methods and) 276.5 490 P
-0.14 (is usually assumed as a given for AI tasks. It often serves as the representation of the asso-) 108 472 P
0.92 (ciated \322world\323 knowledge about the task. In other words, explicit knowledge is a repre-) 108 454 P
0.28 (sentation of the task environment that is internal to the problem solver) 108 436 P
0.28 (. The quality of the) 447.26 436 P
0.21 (problem solving and search methods that are standard in AI are chie\337y determined by the) 108 418 P
0.38 (quality of explicit knowledge they contain. It is no wonder that the phrase \322in the knowl-) 108 400 P
2.04 (edge lies the power) 108 382 P
2.04 (,\323 coined by Edward Feigenbaum, was a popular catch phrase for) 206.56 382 P
(knowledge-based AI through the 1980s.) 108 364 T
0.77 (Every action of an agent using a literal knowledge level model is explicitly account-) 126 334 P
0.63 (able to some component within the agent\325) 108 316 P
0.63 (s knowledge-base. This reduction from knowl-) 312.33 316 P
0.16 (edge-level observations to complete accountability within the representations of the agent) 108 298 P
0.74 (leads to classic problems in arti\336cial intelligence. In order for problem solving ability to) 108 280 P
0.03 (improve on a given task when using knowledge-based AI techniques, the quality of repre-) 108 262 P
1.05 (sentation for the task environment within the problem solver must improve. For a more) 108 244 P
0.12 (accurate problem solver) 108 226 P
0.12 (, more accurate knowledge must be provided. If knowledge of the) 223 226 P
0.63 (task environment is incomplete then the ability of the problem solver is also incomplete.) 108 208 P
1.52 (This direct relationship between quantity of knowledge and quality of ability in strong) 108 190 P
0.99 (methods leads to several classic dif) 108 172 P
0.99 (\336culties in AI: scaling of method, credit assignment,) 281.59 172 P
0.53 (the knowledge acquisition bottleneck, the knowledge indexing problem and the selection) 108 154 P
(of appropriate representations. The next few subsections review each of these problems.) 108 136 T
FMENDPAGE
%%EndPage: "4" 5
%%Page: "5" 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
(5) 534 712 T
108 90 540 702 R
7 X
V
1 F
0 X
(1.2.1  Cr) 108 694 T
(edit Assignment Pr) 151.75 694 T
(oblem) 249.47 694 T
0 F
0.6 (The) 126 668 P
2 F
0.6 (cr) 148.24 668 P
0.6 (edit assignment pr) 157.79 668 P
0.6 (oblem) 247.15 668 P
0 F
0.6 ( \050Minsky 1963\051 highlights the issue of how to convert) 276.47 668 P
0.96 (feedback about problem solving into information about how to manipulate a knowledge) 108 650 P
1.87 (structure internal to the problem solver) 108 632 P
1.87 (. There are two forms of the credit assignment) 303.55 632 P
0.12 (problem: global and local. The) 108 614 P
2 F
0.12 (global cr) 259.18 614 P
0.12 (edit assignment pr) 302.51 614 P
0.12 (oblem) 390.93 614 P
0 F
0.12 ( is the determination that) 420.24 614 P
1.57 (there is in fact an error in the internal structure. T) 108 596 P
1.57 (ypically) 359.63 596 P
1.57 (, explicit goals or evaluation) 397.49 596 P
0.41 (functions within the problem solver determine the global correctness of its internal struc-) 108 578 P
2.15 (ture. In the case of explicit knowledge, global credit assignment is determined by its) 108 560 P
-0.23 (inability to correctly solve the problem at hand. The) 108 542 P
2 F
-0.23 (local cr) 358.46 542 P
-0.23 (edit assignment pr) 394.77 542 P
-0.23 (oblem) 482.49 542 P
0 F
-0.23 ( is the) 511.8 542 P
0.22 (identi\336cation of individual components of the internal structure that are in fact erroneous.) 108 524 P
0.15 (Often, it is the local credit assignment problem that poses the most dif) 108 506 P
0.15 (\336culty in construct-) 445.75 506 P
(ing a problem solver) 108 488 T
(.) 206.27 488 T
-0.2 (Knowledge-based AI\325) 126 458 P
-0.2 (s answer to the local credit assignment problem is to add explicit) 231.38 458 P
0.87 (knowledge to the problem solver that describes how to identify the faulty structure \050e.g.) 108 440 P
-0.23 (W) 108 422 P
-0.23 (inston 1975; Minton et al. 1989; Schank and Leake 1989\051. This knowledge is often task-) 118.84 422 P
-0 (speci\336c and relates the feedback from the task environment directly to faulty components.) 108 404 P
1 F
(1.2.2  The Knowledge Acquisition Bottleneck) 108 368 T
0 F
1.16 (The knowledge acquisition problem is typically cast as the dif) 126 342 P
1.16 (\336culty in determining) 433.73 342 P
-0.05 (how a program should interact with an expert so that the expert\325) 108 324 P
-0.05 (s knowledge can be incor-) 414.29 324 P
0.35 (porated into the problem solver \050Hayes-Roth et al. 1983\051. This creates a bottleneck in the) 108 306 P
1.88 (populating the method with the knowledge required to solve the task. The knowledge) 108 288 P
1.39 (acquisition bottleneck can be generalized to the problem of extracting knowledge from) 108 270 P
0.66 (any task representation external to the problem solver and incorporating it into the prob-) 108 252 P
0.51 (lem solver) 108 234 P
0.51 (. The root of this problem lies in the transduction from the external task repre-) 158.15 234 P
1.35 (sentation into the format of the internal environment. As in the local credit assignment) 108 216 P
0.7 (problem above, knowledge-based AI techniques simply supply explicit knowledge inter-) 108 198 P
0.07 (nal to the problem solver that describes how to acquire knowledge from the given domain) 108 180 P
(\050e.g. Davis 1979\051.) 108 162 T
FMENDPAGE
%%EndPage: "5" 6
%%Page: "6" 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
(6) 534 712 T
108 90 540 702 R
7 X
V
1 F
0 X
(1.2.3  Memory \050Knowledge\051 Indexing Pr) 108 694 T
(oblem) 313.99 694 T
0 F
0.18 (In methods such as Case-Based Reasoning \050CBR\051 \050Schank 1982; Kolodner 1989\051 and) 126 668 P
2.33 (Explanation-Based Learning \050EBL\051 \050De Jong and Mooney 1986; Minton et al. 1989;) 108 650 P
1.2 (Schank and Leake 1989\051 knowledge is stored in memory as \322experience\323 and retrieved) 108 632 P
0.18 (when the situation suggests its appropriateness. This can be generalized to the problem of) 108 614 P
-0.29 (indexing any lar) 108 596 P
-0.29 (ge knowledge-base where only a small percentage of the knowledge is rel-) 185.15 596 P
0.12 (evant at any one time \050Schank 1987\051. T) 108 578 P
0.12 (o search a lar) 297.23 578 P
0.12 (ge knowledge-base each time knowl-) 360.98 578 P
1.82 (edge is required for a problem would be prohibitive. CBR and EBL researchers often) 108 560 P
0.81 (assume an indexing scheme for the knowledge that uses features salient to the task. The) 108 542 P
2 (identi\336cation of these features is task dependent to allow all similar knowledge to be) 108 524 P
-0.08 (retrieved with a speci\336c index. Also, the similarity metric is often task-speci\336c, forcing an) 108 506 P
(experience-base to be reindexed when the task changes.) 108 488 T
1 F
(1.2.4  The Pr) 108 452 T
(oblem of Scaling) 173.41 452 T
0 F
1.97 (A simple relation of AI systems is that as the accuracy of the explicit knowledge) 126 426 P
-0.19 (improves, the accuracy of the problem solver also improves. However) 108 408 P
-0.19 (, to become an order) 442.52 408 P
0.17 (of magnitude more accurate, the quantity of explicit knowledge or processing required by) 108 390 P
0.5 (the problem solver often increases exponentially or worse. Such a discrepancy of scaling) 108 372 P
1.05 (between the accuracy of the problem solver and the algorithmic ef) 108 354 P
1.05 (fort is the chief dif) 436.07 354 P
1.05 (\336-) 529.34 354 P
(culty faced in AI techniques.) 108 336 T
1.13 (For a complex problem the quantity of internal knowledge required to represent the) 126 306 P
0.48 (task environment suf) 108 288 P
0.48 (\336ciently may be prohibitive to extract and/or provide explicitly) 209.34 288 P
0.48 (. The) 514.87 288 P
-0.1 (popularity of blocksworld tasks \050e.g. W) 108 270 P
-0.1 (inograd 1972; W) 297.24 270 P
-0.1 (altz 1975\051 was chie\337y due to their) 377.36 270 P
1.24 (concisely represented task environment. The \336niteness of the situations that could arise) 108 252 P
0.51 (allowed a complete encoding as explicit knowledge. This allows a) 108 234 P
2 F
0.51 (closed-world assump-) 433.89 234 P
1.28 (tion) 108 216 P
0 F
1.28 ( to restrict the knowledge needed to solve the problems. The extreme of the block-) 126.66 216 P
0.74 (sworld approach requires a complete and hence enormous encyclopedic knowledge-base) 108 198 P
1.89 (of \322common sense\323 internal to a general problem solver \050Lenat, Prakash and Shepard) 108 180 P
(1986\051.) 108 162 T
FMENDPAGE
%%EndPage: "6" 7
%%Page: "7" 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
(7) 534 712 T
108 90 540 702 R
7 X
V
1 F
0 X
(1.2.5  Repr) 108 694 T
(esentation Design) 163.74 694 T
0 F
1.81 (The volume of information required to solve some complex problems often forces) 126 668 P
0.73 (researchers to spend a signi\336cant amount of time determining an appropriate representa-) 108 650 P
0.15 (tion that reduces the explicit knowledge to a manageable level. Often this is done to com-) 108 632 P
1.15 (pensate for inadequacies inherent in the technique itself \050e.g. McClelland nd Rumelhart) 108 614 P
3.42 (1986b\051. Given that a particular task has a speci\336c computational ability) 108 596 P
3.42 (, the game) 483.53 596 P
0.21 (becomes can the task be represented in such a way that the limited technique can solve it.) 108 578 P
-0.03 (This has lead to a number of representations to be proposed for problem solving including) 108 560 P
2.03 (production systems \050Newel and Simon 1981\051, predicate calculus \050Nilson 1980\051, fuzzy) 108 542 P
0.66 (logic \050Zadeh 1965\051, connectionist networks \050Rumelhart and McClelland 1986\051, semantic) 108 524 P
1.37 (networks \050Brachman 1979\051, frames \050Minsky 1975\051, conceptual structures \050Sowa 1984\051,) 108 506 P
2.66 (scripts \050Schank and Ableson 1977\051, functional representations \050Sembugamoorthy and) 108 488 P
2.75 (Chandrasekaran 1986\051, the multiple representations of generic tasks \050Chandrasekaran) 108 470 P
-0.28 (1986\051 and others. This representationalist perspective to AI problem solving adheres to the) 108 452 P
(philosophy \322Given the right task representation, magic happens.\323) 108 434 T
0.66 (There is a signi\336cant problem with manipulating the task representation for maximal) 126 404 P
-0.11 (computational bene\336t. The directed design of the problem solver) 108 386 P
-0.11 (\325) 418.32 386 P
-0.11 (s representation assumes) 421.65 386 P
0.04 (knowledge of the task, the algorithm to solve it and how the task is best represented given) 108 368 P
1.47 (the algorithm to solve it. This forces the human programmer to remain in the problem) 108 350 P
1.78 (solving loop in order to determine an adequate representation before problem solving.) 108 332 P
0.19 (However) 108 314 P
0.19 (, it is often said that the true intelligence for solving problem comes in the deter-) 151.48 314 P
0.03 (mination of an appropriate representation for problem solving \050Newell 1981, 1982; Reeke) 108 296 P
(and Edelman 1988\051.) 108 278 T
1 F
(1.3  The Root of the Pr) 108 242 T
(oblems) 224.04 242 T
0 F
0.88 (All of the limitations for AI methods that use explicit knowledge stem from the fact) 126 218 P
2 (that the task environment is represented internally in the problem solver) 108 200 P
2 (. The explicit) 472.39 200 P
0.45 (knowledge, by virtue of its task-speci\336c nature, is a representation or model of the exter-) 108 182 P
-0.07 (nal task environment that is) 108 164 P
2 F
-0.07 (dir) 243.25 164 P
-0.07 (ectly) 256.8 164 P
0 F
-0.07 ( available to the problem solver) 279.44 164 P
-0.07 (. Because this model is) 430 164 P
-0.23 (internal, it is manipulatable, can be rearranged and augmented, and can be employed using) 108 146 P
0.34 (only weak methods if necessary) 108 128 P
0.34 (. But once the commitment to include explicit knowledge) 261.79 128 P
0.83 (is made, then the classic problems of AI arise. Because the task environment is external) 108 110 P
FMENDPAGE
%%EndPage: "7" 8
%%Page: "8" 8
612 792 0 FMBEGINPAGE
108 63 540 702 R
7 X
0 K