## Identifying Structural Mechanism in Standard GP (2003)

Venue: | In Proceedings of GECCO 2003 |

Citations: | 9 - 3 self |

### BibTeX

@INPROCEEDINGS{Daida03identifyingstructural,

author = {Jason M. Daida and Adam M. Hilss},

title = {Identifying Structural Mechanism in Standard GP},

booktitle = {In Proceedings of GECCO 2003},

year = {2003},

pages = {1639--1651},

publisher = {Springer-Verlag}

}

### OpenURL

### Abstract

Abstract. This paper presents a hypothesis about an undiscovered class of mechanisms that exist in standard GP. Rather than being intentionally designed, these mechanisms would be an unintended consequence of using trees as information structures. A model is described that predicts outcomes in GP that would arise solely from such mechanisms. Comparisons with empirical results from GP lend support to the existence of these mechanisms. 1

### Citations

749 |
Genetic Programming II: Automatic Discovery of Reusable Programs
- Koza
- 1994
(Show Context)
Citation Context ...ultiplexer. Furthermore, we placed the outcomes from these problems in the context of both our theory and in the prevailing theory [3, 25]. The setup and implementation of quintic and sextic are from =-=[21]-=-; the multiplexer problems that use arity-2 functions, [22]. The GP parameters used were Koza’s [21]: population 4,000; tournament selection M = 7; maximum generations 50. The maximum depth was set to... |

703 | No free lunch theorems for optimization
- Wolpert, Macready
- 1997
(Show Context)
Citation Context ...s an egalitarian view of representation and evolutionary algorithms: there is no single representation or evolutionary algorithm that is universally superior to the rest when it comes to optimization =-=[7]-=-. However, the matter of trees as information structure is distinct from the matter of trees as an issue of problem representation. When one does an analysis in problem representation, one implicitly ... |

696 | Mersenne Twister: A 623-dimensionally equidistributed uniform pseudorandom number generator. Association for Computer Machinery Transactions Modeling and Computer Simulation (TOMACS
- Matsumoto, Nishimura
(Show Context)
Citation Context ...set to 512. We used a modified version of lilgp [23] as was used in [17]. Most of the modifications were for bug fixes and for the replacement of the random number generator with the Mersenne Twister =-=[24]-=-. We configured lilgp to run as a single thread. Internal node bias was set at default (0.9). Figure 9 shows the results for these test problems. Each dot represents a best-of-trial individual out of ... |

353 |
The Art of Computer Programming, Volume 1, Fundamental Algorithms
- Knuth
- 1973
(Show Context)
Citation Context ...atical entities apart from the information such trees would carry. Consequently, it is not unusual to treat information structure as a level of abstraction below that of problem representation (as in =-=[8]-=-). Whereas there are several problem representations that have been of interest in evolutionary computation (i.e., binary strings, real-valued vectors, finite-state representations, and parse trees), ... |

231 | Foundations of genetic programming
- Langdon, Poli
- 2002
(Show Context)
Citation Context ...d that something else is at work, while Langdon et al. has hypothesized that the causes for what Soule and Foster have found was a consequence of a random walk on the distribution of tree shapes (see =-=[2, 3]-=-). However, according to this paper, the answer would be “yes, additional mechanisms do exist.” We offer a hypothesis that identifies several of these mechanisms, all of which exist at a layer of abst... |

112 |
Diffusion-limited aggregation and kinetical critical phenomena. Phys. Rev.Lett
- Witten, Sander
- 1981
(Show Context)
Citation Context ... shape distribution because GP trees grow iteratively. This seemingly minor distinction is analogous to the distinction in physics between purely diffusive processes and diffusion-limited aggregation =-=[26]-=-. The hypothesis and model have suggested a method of visualizing tree structures in GP. The method was briefly mentioned in this paper and in [28]. A detailed account is given in this proceedings [16... |

102 | Exons and Code Growth in Genetic Programming
- Soule
- 2002
(Show Context)
Citation Context ...nswer might still be “probably not,” although there has been previous work that suggests that there are other considerations that need to be made. The line of investigation by Soule and Foster (e.g., =-=[1]-=-) implied that something else is at work, while Langdon et al. has hypothesized that the causes for what Soule and Foster have found was a consequence of a random walk on the distribution of tree shap... |

100 | The Average Height of Binary Trees and Other Simple Trees - Flajolet, Odlyzko - 1982 |

92 | The Evolution of Size and Shape
- Langdon, Soule, et al.
- 1999
(Show Context)
Citation Context ...d that something else is at work, while Langdon et al. has hypothesized that the causes for what Soule and Foster have found was a consequence of a random walk on the distribution of tree shapes (see =-=[2, 3]-=-). However, according to this paper, the answer would be “yes, additional mechanisms do exist.” We offer a hypothesis that identifies several of these mechanisms, all of which exist at a layer of abst... |

38 | Quadratic Bloat in Genetic Programming
- Langdon
- 2000
(Show Context)
Citation Context ...problems in the context of both our theory and in the prevailing theory [3, 25]. The setup and implementation of quintic and sextic are from [21]; the multiplexer problems that use arity-2 functions, =-=[22]-=-. The GP parameters used were Koza’s [21]: population 4,000; tournament selection M = 7; maximum generations 50. The maximum depth was set to 512. We used a modified version of lilgp [23] as was used ... |

36 | What Makes a Problem GP-Hard? Analysis of a Tunably Difficult - Daida, Bertram, et al. - 2001 |

28 | Random Generation of Trees - Alonso, Schott - 1995 |

19 | What makes a problem GP-hard? validating a hypothesis of structural causes
- Daida, Li, et al.
(Show Context)
Citation Context ...rd Genetic Programming 1649 For 11-input multiplexer, Region I accounts for 92%, in comparison to the prevailing theory (0.62%). Another comparison with empirical results is given in this proceedings =-=[27]-=-. This work introduces a test called the Lid problem, which is anticipated by our hypothesis and model. Indeed, if structure alone accounts for significant GP behavior, it should be possible to devise... |

15 | Size Fair and Homologous Tree Crossovers for Tree Genetic Programming
- Langdon
- 2000
(Show Context)
Citation Context ...r, we considered quintic, sextic, 6-input multiplexer, and 11-input multiplexer. Furthermore, we placed the outcomes from these problems in the context of both our theory and in the prevailing theory =-=[3, 25]-=-. The setup and implementation of quintic and sextic are from [21]; the multiplexer problems that use arity-2 functions, [22]. The GP parameters used were Koza’s [21]: population 4,000; tournament sel... |

11 | A new evolutionary approach to the degree constrained minimum spanning tree problem - Knowles, Corne - 2000 |

9 |
Analysis of single-node (building) blocks in genetic programming
- Daida, Bertram, et al.
- 1999
(Show Context)
Citation Context ...he GP parameters used were Koza’s [21]: population 4,000; tournament selection M = 7; maximum generations 50. The maximum depth was set to 512. We used a modified version of lilgp [23] as was used in =-=[17]-=-. Most of the modifications were for bug fixes and for the replacement of the random number generator with the Mersenne Twister [24]. We configured lilgp to run as a single thread. Internal node bias ... |

9 | Limits to Expression in Genetic Programming: Lattice-Aggregate Modeling
- Daida
(Show Context)
Citation Context ...el. Indeed, if structure alone accounts for significant GP behavior, it should be possible to devise a structure-only problem in GP that demonstrates this. Other empirical comparisons can be found in =-=[28]-=-. 6 Conclusions This paper described a hypothesis of an undiscovered class of mechanisms that exist in standard GP. Rather than being intentionally designed, these structural mechanisms have been hypo... |

8 | A Hybrid GP Approach for Numerically Robust Symbolic Regression - Raidl - 1998 |

6 |
The Collected Mathematical Papers
- Cayley
- 1989
(Show Context)
Citation Context ...rom work in GP. 1 Trees were likely first formalized as a mathematical entity in [13]. The term tree and the beginnings of an extensive treatment of trees as a mathematical entity likely started with =-=[14]-=-. Trees overtly represented information in computer memory in some of the earliest machine-language 1 Langdon has used results from [9-10] in his analyses of GP performance. His incorporation of the m... |

5 | Characterizing a Tunably Difficult Problem in Genetic Programming - Chaudhri - 2000 |

4 |
Guidelines for a Suitable Encoding
- Fogel, Angeline
- 1997
(Show Context)
Citation Context ...r in which one represents an appropriate solution—say, a vector of real values—in turn drives the type of operators, the type of objective function, even the type of evolutionary algorithm to be used =-=[6]-=-. Representations, and their corresponding evolutionary algorithms, are graded according to their worth in solving a particular problem. In the post-NFL era, one takes an egalitarian view of represent... |

2 |
lilgp
- Zongker, Punch
- 1995
(Show Context)
Citation Context ...2 functions, [22]. The GP parameters used were Koza’s [21]: population 4,000; tournament selection M = 7; maximum generations 50. The maximum depth was set to 512. We used a modified version of lilgp =-=[23]-=- as was used in [17]. Most of the modifications were for bug fixes and for the replacement of the random number generator with the Mersenne Twister [24]. We configured lilgp to run as a single thread.... |

2 |
Parse Trees.” Handbook of Evolutionary Computation
- Angeline
- 1997
(Show Context)
Citation Context ...urprisingly, stems from the choice of using trees as an information structure. In evolutionary computation, the issue of trees would typically fall under the category of problem representation (e.g., =-=[4]-=-). One of the basic considerations in evolutionary computation is the matter of choosing a representation that is suited for solving a particular problem [5]. The manner in which one represents an app... |

1 |
Parse Trees," in Handbook of Evolutionary Computation, T. Bäck, et al., Eds
- Angeline
- 1997
(Show Context)
Citation Context ...urprisingly, stems from the choice of using trees as an information structure. In evolutionary computation, the issue of trees would typically fall under the category of problem representation (e.g., =-=[4]-=-). One of the basic considerations in evolutionary computation is the matter of choosing a representation that is suited for solving a particular problem [5]. The manner in which one represents an app... |

1 |
Introduction [to Issues in Representations]," in Handbook of Evolutionary Computation
- Deb
- 1997
(Show Context)
Citation Context ... category of problem representation (e.g., [4]). One of the basic considerations in evolutionary computation is the matter of choosing a representation that is suited for solving a particular problem =-=[5]-=-. The manner in which one represents an appropriate solution—say, a vector of real values—in turn drives the type of operators, the type of objective function, even the type of evolutionary algorithm ... |

1 |
Annalen der Physik und Chemie
- Kirchoff
(Show Context)
Citation Context ... Trees as information structures represent an extensive body of literature that has been, for the most part, separate from work in GP. 1 Trees were likely first formalized as a mathematical entity in =-=[13]-=-. The term tree and the beginnings of an extensive treatment of trees as a mathematical entity likely started with [14]. Trees overtly represented information in computer memory in some of the earlies... |

1 |
presented at Symposium on Automatic Programming
- Kahrimanian
- 1954
(Show Context)
Citation Context ...purpose of algebraic formula manipulation ([8], 459). Because tree structures were well suited for formula manipulation, these structures were used in some of the 1950’s work in automatic programming =-=[15]-=-. 2.2 Visualizing Tree Structures To appreciate the role of tree structure in forming solutions in GP, it helps to visualize the kinds of structures that occur. Visualizing structure, particularly for... |

1 |
Hilss,"Visualizing Tree Structures in Genetic Programming," GECCO
- Daida, A
- 2003
(Show Context)
Citation Context ...mapped to labels of a full binary tree by showing that a set of labels corresponding to A is a subset of labels corresponding to C. (Note: a formal description of our visualization method is found in =-=[16]-=-.) Figure 2 shows an example of a full binary tree of depth 10 (2047 nodes) and an example of binary tree of depth 26 (461 nodes) from GP. From the standpoint of typical measures of tree structure (i.... |

1 |
Introduction [to Issues in Representations]," in Handbook of Evolutionary Computation
- Deb
- 1997
(Show Context)
Citation Context ... category of problem representation (e.g., [4]). One of the basic considerations in evolutionary computation is the matter of choosing a representation that is suited for solving a particular problem =-=[5]-=-. The manner in which one represents an appropriate solution—say, a vector of real values—in turn drives the type of operators, the type of objective function, even the type of evolutionary algorithm ... |

1 |
Hilss,"Visualizing Tree Structures in Genetic Programming," GECCO
- Daida, A
- 2003
(Show Context)
Citation Context ...mapped to labels of a full binary tree by showing that a set of labels corresponding to A is a subset of labels corresponding to C. (Note: a formal description of our visualization method is found in =-=[16]-=-.) Figure 2 shows an example of a full binary tree of depth 10 (2047 nodes) and an example of binary tree of depth 26 (461 nodes) from GP. From the standpoint of typical measures of tree structure (i.... |