J. R. Koza, Genetic programming, On the programming of computers by means of natural selection, MIT Press, Cambridge, MA, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Unfortunately, learning procedures always use the same strategy to adapt the weights, disregarding the fact that the error surfaces associated with different tasks present completely different features. Attempts to use genetic programming to evolve the architecture and the weights simultane ously [4, 5] have been marred by the fact that parse trees are not suitable for representing oriented graphs. Alternatively, GP has been used to evolve rules for constructing neural networks [6, 7, 8] However, this approach imposes constraints on the weights of the neural network. In this paper, a new ....

....second parent. To prevent the excessive growth of the parse trees, a maximum depth is specified, and the selection of the subtrees for swapping is carried out according to this constraint. A bias is also introduced to favor functions in the se lection of the roots of the subtrees for crossover [4]. In addition, two terminals representing the variables are never selected as subtrees for crossover, since they encode the same weight. 4 Experimental results In the experiments described in this section, all indi viduals were initialized with 10 internal nodes in a sin gle internal layer. ....

[Article contains additional citation context not shown here]

J. Koza. Genetic Programming, on the Program- ming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....using 2 Gamma1 1 control line multiplexers. Any implementation using less than that number of elements could be considered an improvement in the design. Since the optimal minimum number needed is unknown for most of the logic functions, the use of a heuristic such as genetic programming [7] seems adequate. The implementation cost measured in terms of silicon surface has been studied for many years. Consider an n variable multiplexer realized by means of 2 multiplexers with 1 control signal. Assuming that the cost of a single unit is K, the cost of such a realization is ....

....the control signal of the multiplexer. 3. Previous work It is possible to find in the literature several reports concerning the design of combinational logic circuits using GAs. Louis [11] was one of the first researchers who reported this class of work. Further work has been reported by Koza [7], Coello et al. 2, 3] Iba et al. 6] and Miller et al. 12] However, none of these approaches has concentrated on the exclusive use of multiplexers to design combinational circuits using evolutionary techniques, although It is worth mentioning that Koza s approach to the design of ....

J. R. Koza. Genetic Programming. On the Programming of Computers by Means of Natural Selection. The MIT Press, 1992.

....introns [Angeline98] which are code segments that are executed but have no e#ect on the overall result. For instance, the codes ( 0 a) and (not (not x) include the semantic introns. These introns can be edited and replaced using a predefined template, such as double not s or zero plus [Koza 92] Preliminary experiments have shown that editing semantic introns did not any harm or good to our multi agent GP learning. This is partly due to the fact that these introns seldom occur in our task. We will be in pursuit of the role of these introns in our future research. 6 Conclusion This ....

Koza, J., Genetic Programming, On the Programming of Computers by means of Natural Selection, MIT Press, 1992

....It could later be used to relocate the robot if odometry information becomes inaccurate or is completely absent due to a system malfunction. 4 Symbolic regression using genetic programming To evolve the inverse mapping from sensor readings to robot localizations we are using genetic programming [8, 10, 2]. Koza [9] has shown that genetic programming can effectively be used to search for a function described by a finite number of mappings. To apply genetic programming to the task of robot localization we first have to define the set of input variables. The task of finding a function which maps raw ....

J. R. Koza. Genetic Programming, On the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusetts, 1992.

....Unfortunately, learning procedures always use the same strategy to adapt the weights, disregarding the fact that the error surfaces associated with different tasks present completely different features. Attempts to use genetic programming to evolve the architecture and the weights simultaneously [4, 5] have been marred by the fact that parse trees are not suitable for representing oriented graphs. Alternatively, GP has been used to evolve rules for constructing neural networks [6, 7, 8] However, this approach imposes constraints on the weights of the neural network. In this paper, a new ....

....the second parent. To prevent the excessive growth of the parse trees, a maximum depth is specified, and the selection of the subtrees for swapping is carried out according to this constraint. A bias is also introduced to favor functions in the selection of the roots of the subtrees for crossover [4]. In addition, two terminals representing the variables were never selected as subtrees for crossover, since they encode the same weight. 4 Experimental results In the experiments described in this section, all individuals were initialized with 10 internal nodes in a single internal layer. ....

[Article contains additional citation context not shown here]

J. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992. 12

....naturally be represented as vectors. They are oriented graphs, whose nodes are neurons and whose arcs are synaptic connections. Therefore, it is arguable that any efficient approach to evolve ANNs should use operators based on this structure. Some recent work based on genetic programming (GP) [21], originally developed to evolve computer programs, is a first step in this direction. For example, in [21, 22] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of ....

....synaptic connections. Therefore, it is arguable that any efficient approach to evolve ANNs should use operators based on this structure. Some recent work based on genetic programming (GP) 21] originally developed to evolve computer programs, is a first step in this direction. For example, in [21, 22] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly with parse trees either. Indeed, an alternative approach ....

[Article contains additional citation context not shown here]

J. R. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....be represented as binary vectors. They are oriented graphs, whose nodes are neurons and whose arcs are synaptic connections. Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on Genetic Programming (GP) [10], a special form of GA, originally developed to evolve computer programs, is a first step in this direction. For example, in [10, 11] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the ....

....Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on Genetic Programming (GP) 10] a special form of GA, originally developed to evolve computer programs, is a first step in this direction. For example, in [10, 11] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly as parse trees either. Indeed, an alternative approach based ....

[Article contains additional citation context not shown here]

J. R. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....introns [Angeline98] which are code segments that are executed but have no e#ect on the overall result. For instance, the codes ( 0 a) and (not (not x) include the semantic introns. These introns can be edited and replaced using a predefined template, such as double not s or zero plus [Koza 92] Preliminary experiments have shown that editing semantic introns did not any harm or good to our multi agent GP learning. This is partly due to the fact that these introns seldom occur in our task. We will be in pursuit of the role of these introns in our future research. 6 Conclusion This ....

Koza, J., Genetic Programming, On the Programming of Computers by means of Natural Selection, MIT Press, 1992

....be represented as binary vectors. They are oriented graphs, whose nodes are neurons and whose arcs are 1 synaptic connections. Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on genetic programming (GP) [10], originally developed to evolve computer programs, is a first step in this direction. For example, in [10, 11] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of ....

....1 synaptic connections. Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on genetic programming (GP) 10] originally developed to evolve computer programs, is a first step in this direction. For example, in [10, 11] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly with parse trees either. Indeed, an alternative approach ....

[Article contains additional citation context not shown here]

J. R. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

.... . J I K1 MQ 9 k 3 H rL ;X 7 F k # 2 1 J1 7 K KGP N a 8 J W m 0 i d m C H N J2= 9 kMM;R K r ( 7 # 3 N iJ, k h K GP G O W m 0 i r J2= 5 ; FL E H 9 k m C H N )8f d9=B J N 7W r;n F k # GP N4pK E ;WA[ OStanfordBg3X NJohn Koza i K h jDs0F 5 l [Koza92] # GP OGA K F0lJ,Ln r3NN) 7 D D k #GP N9M ( rAI KE,MQ 73X= dO LdBj2r7h r B8= 9 k;n r J2=O E 3X= Jevolutionary learning K H8F V # 3 l OI=8= 5 l CN 1 rJQ49 7 A BrEqBA K h jE, Z J2r r;D 7 F E,9gE J3X= jK G k # J2=O E 3X= O i 7 U 7 9 F J I ....

.... DL o NO M 5 9f rMQ k H f(D 2 , D 1 , D 0 ) D 2 D 1 D 0 # D 2 D 1 D 0 # D 2 D 1 D 0 # D 2 D 1 D 0 (2) H=q 1 k # 7 x # y Ox Hy NO M OB xy Ox Hy NO M Q x Ox N H]Dj rI= 9 # 3 N4X t rF NO H=PNO NBP iGP G3X= 7 F h [Koza92] #GP N Q i a H 7 F O N b N r:NMQ 9 k # 1. F = AND 2 , OR 2 , NOR 2 , NOR 2 Hs= C 5 9f ODL o N V k4X t G k # 72 IU N t;z O0z t N t rI= 9 # 2. T = D0, D1, D2 = C 5 9f O V k4X t NF NOJQ t G k # 8 3. E,9gEY O 9 Y F NF NO Q s(2 3 8D) ....

[Article contains additional citation context not shown here]

Koza, J. Genetic Programming, On the Programming of Computers by means of Natural Selection, MIT Press, 1992

....that our approach is superior to usual neural networks in terms of generalization of learning. 1 Introduction Most of Genetic Programming (GP) 1 computation time is taken up in measuring the fitness of GP trees. Naturally, the amount of time depends directly upon the size of the trees [Koza92] The size of GP trees is usually controlled by user defined parameters, such as the maximum number of nodes or maximum tree depth. We tried Koza style GP to obtain a solution tree to the 6 multiplexer problem [Higuchi93] see equation (1) using the function set [AND, OR, IF, NOT] and using ....

.... To overcome this difficulty and to get better results with much lesser population, we introduce a Minimum Description Length (MDL) principle [Rissanen78] to define fitness functions 1 Note that throughout this paper, the term Genetic Programming (GP) is used according to the Koza definition [Koza92] and not according to the deGaris definition [deGaris93] Unfortunately, there are two definitions of GP in the literature. 1 in GP, so as to control the growth of trees. We choose a decision tree representation for the chromosomes, and show how an MDL principle can be used to define the GP ....

Koza, J. Genetic Programming, On the Programming of Computers by means of Natural Selection, MIT Press, 1992

....introns [Angeline98] which are code segments that are executed but have no e#ect on the overall result. For instance, the codes ( 0 a) and (not (not x) include the semantic introns. These introns can be edited and replaced using a predefined template, such as double not s or zero plus [Koza 92] Preliminary experiments have shown that editing semantic introns did not any harm or good to our multiagent GP learning. This is partly due to the fact that these introns seldom occur in our task. We will be in pursuit of the role of these introns in our future research. As a future research ....

Koza, J., Genetic Programming, On the Programming of Computers by means of Natural Selection, MIT Press, 1992

....GA as a design tool. Unfortunately, the incorporation of knowledge into the GA decreases its usefulness as a general search tool. Louis overcomes this problem by defining an operator that he claims to be domain independent, but whose efficiency turns out to depend on the representation used. Koza [21] has used genetic programming to design combinational circuits. He has designed, for example, a two bit adder, using a small set of gates (AND, OR, NOT) but his emphasis has been on generating functional circuits rather than on optimizing them. In fact, this is also the case in Louis research, ....

....more powerful for structural design in general [33] However, genetic programming produces circuits that are highly redundant and difficult to simplify automatically. Furthermore, the computer resources normally required to produce such circuits are very demanding in terms of memory and CPU time [21]. That is why we decided to use instead a matrix representation that is encoded linearly in a chromosome, and turns out to be a compromise between the powerful tree representation used by genetic programming and the relatively limited linear representation used by a conventional genetic algorithm ....

[Article contains additional citation context not shown here]

John R. Koza. Genetic Programming. On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, Massachusetts, 1992.

....it is arguable that any efficient approach to evolve NNs should use operators based on this structure. The authors are with the School of Computer Science, University of Birmingham, Birmingham B15 2TT, UK. E mail: fJ.Pujol,R.Polig cs.bham.ac. uk Some recent work based on genetic programming (GP) [12], originally developed to evolve computer programs, is a first step in this direction. For example, in [12] 13] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of ....

....authors are with the School of Computer Science, University of Birmingham, Birmingham B15 2TT, UK. E mail: fJ.Pujol,R.Polig cs.bham.ac.uk Some recent work based on genetic programming (GP) 12] originally developed to evolve computer programs, is a first step in this direction. For example, in [12], 13] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly with parse trees either. Indeed, an alternative ....

[Article contains additional citation context not shown here]

J. R. Koza, Genetic Programming, on the Programming of Computers by Means of Natural Selection. Cambridge, Massachusets: The MIT Press, 1992.

....graphs) that have to be already fully functional. They also exhibit a lack of generality since the added knowledge reduces their applicability to a certain specific problem. The genetic programming (GP) approach we are to describe, follows the automatic programming capacity proclaimed by Koza [10]. That is, GP synthesizes programs or functions that reproduce a desired behavior. In our system, GP constructs Boolean functions by combining samples taken from the space of partial solutions. Once a 100 functional solution is found, our goal is turned to their minimization. Thus, the fitness ....

....we substituted gates by binary multiplexers. Binary multiplexers are universal function generators (defined later) Thus, they form a sound basis for the synthesis of logic functions. The working hypothesis is that GP can synthesize logic circuits by means of binary multiplexers (muxes, for short) [10], and that the replication of only one element (instead of five or six different gates) will decrease the manufacturing process cost (in this paper we address only the first issue) We emphasized the importance of replication by allowing the use of only 1 control line multiplexers in the ....

[Article contains additional citation context not shown here]

John R. Koza. Genetic Programming. On the Programming of Computers by Means of Natural Selection. The MIT Press, 1992.

....naturally be represented as vectors. They are oriented graphs, whose nodes are neurons and whose arcs are synaptic connections. Therefore, it is arguable that any efficient approach to evolve ANNs should use operators based on this structure. Some recent work based on genetic programming (GP) [21], originally developed to evolve computer programs, is a first step in this direction. For example, in [21, 22] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of ....

....synaptic connections. Therefore, it is arguable that any efficient approach to evolve ANNs should use operators based on this structure. Some recent work based on genetic programming (GP) 21] originally developed to evolve computer programs, is a first step in this direction. For example, in [21, 22] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly with parse trees either. Indeed, an alternative approach ....

[Article contains additional citation context not shown here]

J. R. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....Unfortunately, learning procedures always use the same strategy to adapt the weights, disregarding the fact that the error surfaces associated with different tasks present completely different features. Attempts to use genetic programming to evolve the architecture and the weights simultaneously [4, 5] have been marred by the fact that parse trees are not suitable for representing oriented graphs. Alternatively, GP has been used to evolve rules for constructing neural networks [6, 7, 8] However, this approach imposes constraints on the weights of the neural network. In this paper, a new ....

....the second parent. To prevent the excessive growth of the parse trees, a maximum depth is specified, and the selection of the subtrees for swapping is carried out according to this constraint. A bias is also introduced to favor functions in the selection of the roots of the subtrees for crossover [4]. In addition, two terminals representing the variables are never selected as subtrees for crossover, since they encode the same weight. 4 Experimental results In the experiments described in this section, all individuals were initialized with 10 internal nodes in a single internal layer. ....

[Article contains additional citation context not shown here]

J. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....be represented as binary vectors. They are oriented graphs, whose nodes are neurons and whose arcs are synaptic connections. Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on genetic programming(GP) [8] is a first step in this direction. For example, in [8, 9] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly ....

....graphs, whose nodes are neurons and whose arcs are synaptic connections. Therefore, it is arguable that any efficient approach to evolve NNs should use operators based on this structure. Some recent work based on genetic programming(GP) 8] is a first step in this direction. For example, in [8, 9] neural networks have been represented as parse trees which are recombined using a crossover operator which swaps subtrees representing subnetworks. However, the graph like structure of neural networks is not ideally represented directly with parse trees either. Indeed, an alternative approach ....

[Article contains additional citation context not shown here]

J. R. Koza. Genetic Programming, on the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusets, 1992.

....Lastly, the question of whether GP can scale can only be answered in full by analysing experimental data, but initial indications show that GP can create programs to solve complex problems in other domains. 2 Applying GP to Service Creation 2. 1 Functions and Terminals Classical tree based GP [16] requires a set of functions which form the non leaf nodes in the tree and terminals which form the leaf nodes of the tree. The set of functions and terminals must satisfy the closure and sufficiency properties. Terminals may be side affecting or yield data. For this work, the functions were ....

....that it would be easier to analyse the operation of the evolving programs. 2.2 Achieving Closure Several methods have been proposed to ensure that the closure property is maintained during initial creation and subsequent reproduction. This may be achieved in a number of ways. Firstly Koza [16] restricts the types of arguments and functions to compatible types. For instance, all floating point types as in the symbolic regression examples or logical in the Boolean examples. For simple problems with single data types this is sufficient. Secondly, in strongly typed approaches such as ....

Koza, John, R. Genetic Programming, On the Programming of Computers by Means of Natural Selection. 1 Ed. MIT Press 1992.

No context found.

J. R. Koza, Genetic programming, On the programming of computers by means of natural selection, MIT Press, Cambridge, MA, 1992.

No context found.

J. R. Koza, Genetic programming, On the programming of computers by means of natural selection, MIT Press, Cambridge, MA, 1992.

No context found.

Koza, J.R.: Genetic Programming; on the programming of computers by means of natural selection. MIT Press (1992)

No context found.

J.R. Koza, Genetic Programming, on the programming of computers by means of natural selection, MIT Press, 1992. 165

No context found.

Koza, J.: Genetic programming. On the programming of computers by means of natural selection, The MIT Press (1992).

No context found.

J. R. Koza. Genetic Programming. On the Programming of Computers by Means of Natural Selection. The MIT Press, Cambridge, Massachusetts, 1992.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC