## Density-adaptive learning and forgetting (1993)

### Cached

### Download Links

- [www.ircs.upenn.edu]
- [ftp.cis.upenn.edu]
- [ftp.cis.upenn.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | In Proceedings of the Tenth International Conference on Machine Learning |

Citations: | 22 - 2 self |

### BibTeX

@INPROCEEDINGS{Salganicoff93density-adaptivelearning,

author = {Marcos Salganicoff and Marcos Salganico},

title = {Density-adaptive learning and forgetting},

booktitle = {In Proceedings of the Tenth International Conference on Machine Learning},

year = {1993},

pages = {276--283},

publisher = {Morgan Kaufmann}

}

### Years of Citing Articles

### OpenURL

### Abstract

We describe a density-adaptive reinforcement learning and a density-adaptive forgetting algorithm. This learning algorithm uses hybrid k-D/2k-trees to allow foravariable resolution partitioning and labelling of the input space. The density adaptive forgetting algorithm deletes observations from the learning set depending on whether subsequent evidence is available in a local region of the parameter space. The algorithms are demonstrated in a simulation for learning feasible robotic grasp approach directions and orientations and then adapting to subsequent mechanical failures in the gripper. 1

### Citations

4172 | Classification and Regression Trees - Breiman, Freidman, et al. - 1984 |

1210 |
The Design and Analysis of Spatial Data Structures
- Samet
- 1990
(Show Context)
Citation Context ...aved tree-building schedule may be used, so that the preceding decision-tree is used on-line while its successor is being built either in a background process, or on another processor. Alternatively, =-=[17]-=- discusses some issues in incrementally updating adaptive k-D-trees. This issue might be sidestepped by using the forgetting technique together with other incremental techniques such as ID-5 [21], or ... |

1101 | Instance-based learning algorithms
- Aha, Kibler, et al.
- 1991
(Show Context)
Citation Context ...stepped by using the forgetting technique together with other incremental techniques such as ID-5 [21], or using it with nearest-neighbor prediction techniques that do not form explicit decisiontrees =-=[1]-=-. Acknowledgements The work was supported by an NSF Postdoctoral Research Associateship in Computational Science and Engineering (NSF-CDA92-11136), Navy GrantN001488-K-0630, AFOSR Grants 88-0244, AFOS... |

619 | An algorithm for finding best matches in logarithmic expected time - Friedman, Bentley, et al. - 1977 |

340 | Automatic Programming of Behavior-based Robots using Reinforcement Learning
- Mahadevan, Connell
- 1992
(Show Context)
Citation Context ...echnique to delete obsolete observations using exponential weightdecay based on a nearest neighbor criteria. The approach usesadecaycoe cient to decrement anexperience's weighting, similar to that of =-=[8, 11, 9]-=-. However, this coe cient is a function of the similarity of that given experience to subsequent experiences, rather than a xed value. The weight of an exemplar is decayed and deleted when it goes bel... |

323 | Learning efficient classification procedures and their application to chess endgames - Quinlan - 1983 |

314 |
Learning in Embedded Systems
- Kaelbling
- 1993
(Show Context)
Citation Context ...Since we are interested in nding regions of the attribute space having a pessimistic probability ofsuccess greater than some pmin, we use the outcome ratios to compute a probability interval estimate =-=[6]-=- for the underlying probability of receiving a success in that leaf. Two thresholds, pmin and pmax, are required for splitting. If the lower bound of the probability interval is above the pmin thresho... |

210 | Learning to Coordinate Behaviors
- Maes, Brooks
- 1990
(Show Context)
Citation Context ...echnique to delete obsolete observations using exponential weightdecay based on a nearest neighbor criteria. The approach usesadecaycoe cient to decrement anexperience's weighting, similar to that of =-=[8, 11, 9]-=-. However, this coe cient is a function of the similarity of that given experience to subsequent experiences, rather than a xed value. The weight of an exemplar is decayed and deleted when it goes bel... |

195 |
Classi - cation and regression trees
- Breiman, Friedman, et al.
- 1984
(Show Context)
Citation Context ...ibute space, de ned as the number of exemplars per unit volume of attribute space. The DARLING algorithm takes inspiration from decision tree approaches embodied in Classi cation and Regression Trees =-=[3]-=- and ID-3 [14], along with the geometric learning approaches described by Omohundro [13]. The goal of the algorithm is to identify regions of the input attribute space having a lowerbound estimated pr... |

154 |
Neural Computation and Self-Organizing Maps: an introduction
- Ritter, Martinetz, et al.
- 1992
(Show Context)
Citation Context ... ID-3 with 3-D objects. Bennett [2] worked in robotic grasping of polygonal 2-D puzzle piece task using explanation-based learning and domain theories about uncertainty and grasping. Mel [10], Ritter =-=[15]-=- and Cooperstock have used ; neural-networks, self-organizing feature maps and backpropagation, respectively, for learning visually-guided control of robot arms for grasping. 2 Learning Algorithm Firs... |

86 | Efficient algorithms with neural network behavior - Omohundro - 1987 |

72 | ID5: An Incremental ID3 - Utgoff - 1988 |

56 |
Beyond incremental processing: Tracking concept drift
- Schlimmer, Granger
- 1986
(Show Context)
Citation Context ...pace where the lower-bound probability of succeeding (receiving immediate reward) subject to a 1; con dence value is above some minimum probability required for the task. Also, as noted previously by =-=[18, 7, 11]-=-, an important assumption taken by many learning methods is that the concept (e.g. the environment and task) to be learned is stationary over time. By stationary, we mean that the true underlying proc... |

37 |
Learning time-varying concepts
- Kuh, Petsche, et al.
- 1991
(Show Context)
Citation Context ...pace where the lower-bound probability of succeeding (receiving immediate reward) subject to a 1; con dence value is above some minimum probability required for the task. Also, as noted previously by =-=[18, 7, 11]-=-, an important assumption taken by many learning methods is that the concept (e.g. the environment and task) to be learned is stationary over time. By stationary, we mean that the true underlying proc... |

30 | An Incremental Method for Finding Multivariate Splits for Decision Trees - Utgoff, Brodley - 1990 |

29 |
An algorithm for nding best matches in logarithmic expected time
- Friedman, Bentley, et al.
- 1977
(Show Context)
Citation Context ...pproximates those regions. We desire that the tree approximate those regions of the parameter space with minimum over- and under-estimation. 2.1 Density-Adaptation The algorithm rst builds a k-D-tree =-=[5]-=- based on the distribution of the exemplars in the parameter space, ignoring the outcome labels of each exemplar. This step adaptively partitions the exemplar set into a set of bins each with a roughl... |

27 |
Connectionist Robot Motion Planning: A Neu- rally Inspired Approach to Visually Guided Reaching
- Mel
- 1990
(Show Context)
Citation Context ... extension of ID-3 with 3-D objects. Bennett [2] worked in robotic grasping of polygonal 2-D puzzle piece task using explanation-based learning and domain theories about uncertainty and grasping. Mel =-=[10]-=-, Ritter [15] and Cooperstock have used ; neural-networks, self-organizing feature maps and backpropagation, respectively, for learning visually-guided control of robot arms for grasping. 2 Learning A... |

19 |
Shape Recovery and Segmentation with Deformable Part Models
- Solina
- 1987
(Show Context)
Citation Context ...nstrate the utility of these learning and forgetting algorithms in simulation for the assessment of a robotic grasp's suitability toanobjectwithagiven parametric superellipsoid attribute description =-=[19]-=- and pose. However, the method can be applied to any situation where there are a xed set of actions to evaluate, a reward and a real-valued input attribute space. This work di ers from previous e orts... |

18 |
Learning E cient Classi cation Procedures and their Application to Chess End Games
- Quinlan
- 1983
(Show Context)
Citation Context ...de ned as the number of exemplars per unit volume of attribute space. The DARLING algorithm takes inspiration from decision tree approaches embodied in Classi cation and Regression Trees [3] and ID-3 =-=[14]-=-, along with the geometric learning approaches described by Omohundro [13]. The goal of the algorithm is to identify regions of the input attribute space having a lowerbound estimated probability, p;,... |

17 |
robust adaptive control by learning only forward models
- Fast
- 1992
(Show Context)
Citation Context ...pace where the lower-bound probability of succeeding (receiving immediate reward) subject to a 1; con dence value is above some minimum probability required for the task. Also, as noted previously by =-=[18, 7, 11]-=-, an important assumption taken by many learning methods is that the concept (e.g. the environment and task) to be learned is stationary over time. By stationary, we mean that the true underlying proc... |

17 |
Csl: A cost-sensitive learning system for sensing and grasping objects
- Tan, Schlimmer
- 1993
(Show Context)
Citation Context ...rning methods employed. Dunn [4] employed a two dimensional polygonal representation, and a random search for successful grasps during learning, followed by a 2-D model matching during execution. Tan =-=[20]-=- employed a feature-based sonar depth representation and a cost sensitive extension of ID-3 with 3-D objects. Bennett [2] worked in robotic grasping of polygonal 2-D puzzle piece task using explanatio... |

16 | Geometric learning algorithms
- Omohundro
- 1990
(Show Context)
Citation Context ...DARLING algorithm takes inspiration from decision tree approaches embodied in Classi cation and Regression Trees [3] and ID-3 [14], along with the geometric learning approaches described by Omohundro =-=[13]-=-. The goal of the algorithm is to identify regions of the input attribute space having a lowerbound estimated probability, p;, of succeeding (receiving reward) that is greater than some speci ed minim... |

7 | Automatic discovery of robotic grasp configurations - Dunn, Segen - 1988 |

7 |
ID-5: An incremental ID-3
- Utgo
- 1988
(Show Context)
Citation Context ...ely, [17] discusses some issues in incrementally updating adaptive k-D-trees. This issue might be sidestepped by using the forgetting technique together with other incremental techniques such as ID-5 =-=[21]-=-, or using it with nearest-neighbor prediction techniques that do not form explicit decisiontrees [1]. Acknowledgements The work was supported by an NSF Postdoctoral Research Associateship in Computat... |

5 | Learning and Forgetting for Perception-Action: A Projection-Pursuit and Density-Adaptive Approach - Salganicoff - 1992 |

4 |
An incremental method for nding multivariate splits for decision trees
- Utgo, Brodley
- 1990
(Show Context)
Citation Context ... inductive bias. There are a number of immediate enhancements that can be made to the learning algorithm. The rst is to embed decision trees that use non-axis parallel splits such as perceptron trees =-=[22]-=-, or linear discriminants that might be more e ective, since they will in general provide much better performance for shallow trees. Some simple branch-and-bound tests can be implemented to prevent un... |

2 |
Planning to address uncertainty: An incremental approach employing learning through experience
- Bennett
- 1991
(Show Context)
Citation Context ...rasps during learning, followed by a 2-D model matching during execution. Tan [20] employed a feature-based sonar depth representation and a cost sensitive extension of ID-3 with 3-D objects. Bennett =-=[2]-=- worked in robotic grasping of polygonal 2-D puzzle piece task using explanation-based learning and domain theories about uncertainty and grasping. Mel [10], Ritter [15] and Cooperstock have used ; ne... |

1 |
Automatic discovery of robotic grasp con guration
- Dunn, Segen
- 1988
(Show Context)
Citation Context ...l-valued input attribute space. This work di ers from previous e orts in learning for robotic grasping in terms of action and perceptual representation, as well as the learning methods employed. Dunn =-=[4]-=- employed a two dimensional polygonal representation, and a random search for successful grasps during learning, followed by a 2-D model matching during execution. Tan [20] employed a feature-based so... |

1 |
Learning and Forgetting for Perception-Action: A Projection-Pursuit and Density-Adaptive Approach
- Salganico
- 1992
(Show Context)
Citation Context ...ation, respectively, for learning visually-guided control of robot arms for grasping. 2 Learning Algorithm First, we describe the learning algorithm, Density Adaptive reinforcement learning (DARLING) =-=[16]-=-. Here the term Density refers to the local density ofobservations in the attribute space, de ned as the number of exemplars per unit volume of attribute space. The DARLING algorithm takes inspiration... |