Clustering is often used for discovering structure in data. Clustering systems differ in the objective function used to evaluate clustering quality and the control strategy used to search the space of clusterings. Ideally, the search strategy should consistently construct clusterings of high quality, but be computationally inexpensive as well. In general, we cannot have it both ways, but we can partition the search so that a system inexpensively constructs a `tentative' clustering for initial examination, followed by iterative optimization, which continues to search in background for improved clusterings. Given this motivation, we evaluate an inexpensive strategy for creating initial clusterings, coupled with several control strategies for iterative optimization, each of which repeatedly modifies an initial clustering in search of a better one. One of these methods appears novel as an iterative optimization strategy in clustering contexts. Once a clustering has been construct...

ions are made over the structural information (relations)

plans are probabilistic summaries of specific plans, containing pointers to their components -- abstract states, operators, and subplans -- along with associated probabilities. For example, a generic plan for picking up an object (a manipulation plan) might have three subproblems, analogous to the event described above. Icarus uses the same approach to store route knowledge (navigation plans), with places acting as states and with operators like move and turn. Components of Icarus Our designs for the Icarus architecture call for three main components: a perceptual system (Argus), a planning system (Daedalus), and an execution system (Maeander). Argus and Daedalus invoke the memory system (Labyrinth) to retrieve structured experiences from long-term memory, which include objects, states, and plans. 1 Labyrinth first sorts each component of an experience through memory, starting at the root node of the memory hierarchy. At each level, the memory system uses an evaluation function ca...

The data exploration task can be divided into three interrelated subtasks: (i) feature selection, (ii) discovery, and (iii) interpretation. This paper describes an unsupervised discovery method with biases geared toward partitioning objects into clusters that improve interpretability. The algorithm, ITERATE, employs: (i) a data ordering scheme and (ii) an iterative redistribution operator to produce maximally cohesive and distinct clusters. Cohesion or intra-class similarity is measured in terms of the match between individual objects and their assigned cluster prototype. Distinctness or inter-class dissimilarity is measured by an average of the variance of the distribution matchbetween clusters. We demonstrate that interpretability, from a problem solving viewpoint, is addressed by theintra- and interclass measures. Empirical results demonstrate the properties of the discovery algorithm, and its applications to problem solving.

Abstract not found

Fisher's Cobweb provided a well-defined framework for research on the unsupervised induction of probabilistic concept hierarchies. The system also sparked the development of many successors that extended this framework along various dimensions. In this paper, we summarize the assumptions that Cobweb embodies about the representation, organization, use, and formation of probabilistic concepts, along with experimental studies that examine its sources of power. After this, we consider three systems -- Arachne, Twilix, and Oxbow -- that incorporate significant extensions and present empirical evidence that these improve behavior. In closing, we discuss other paradigms for the unsupervised learning of probabilistic knowledge and their relation to the Cobweb framework. We thank our collaborators, including John Gennari, Kathleen McKusick, Kevin Thompson, and John Allen, for their contributions to the research described in this paper. Grant MDA 903-85-C0324 from the Army Research Insitute sup...

In concept learning, objects in a domain are grouped together based on similarity as determined by the attributes used to describe them. Existing concept learners require that this set of attributes be known in advance and presented in entirety before learning begins. Additionally, most systems do not possess mechanisms for altering the attribute set after concepts have been learned. Consequently, a veridical attribute set relevant to the task for which the concepts are to be used must be supplied at the onset of learning, and in turn, the usefulness of the concepts is limited to the task for which the attributes were originally selected. In order to efficiently accommodate changing contexts, a concept learner must be able to alter the set of descriptors without discarding its prior knowledge of the domain. We introduce the notion of attribute-incrementation, the dynamic modification of the attribute set used to describe instances in a problem domain. We have implemented the capability...

Models of Category Learning Pat Langley 1 (langley@isle.org) Institute for the Study of Learning and Expertise 2164 Staunton Court, Palo Alto, CA 94306 USA Abstract In this paper, we compare the rhetoric that sometimes appears in the literature on computational models of category learning with the growing evidence that different theoretical paradigms typically produce similar results. In response, we suggest that concrete computational models, which currently dominate the field, may be less useful than simulations that operate at a more abstract level. We illustrate this point with an abstract simulation that explains a challenging phenomenon in the area of category learning -- the effect of consistent contrasts -- and we conclude with some general observations about such abstract models. Introduction and Overview Learning is one of the ubiquitous aspects of human behavior, so it seems natural that the process of learning has drawn significant attention within both ...

Machine learning cannot occur without some means to represent the learned knowledge. Researchers have long recognized the influence of representational choices, and the major paradigms in machine learning are organized not around induction algorithms or performance elements as much as around representational classes. Major examples include logical

j_ REPORT DOCUMENTATION PAGE OMBNo.07o4o188 Oijt31!C _e_3rt,r1 _ burden +c,r this, oile(tlOtl]f,nformatlOn,s estimated to average 1 hour per resooqse irlcIu_4_g rife t_me tor re_e'hmg irlstruGIOr;$, sear(rang e,rsDng dat _ sour¢_., _ather_r _:_r_d-na!_t_l_r_g the _a[a needed. 3nd Como/etlng 3nO reviewing lhe c3t{_,<_l©rl,)f rr f,_rma_l©n S_'nd:ommertt _ re_arctmg this burden estbma_e or _rly other asDe _ of l_s c,_lecliQt _ 3f mtormat_O r_. mc)udlrlg sugg_tl_n _, f_r r@d_ng thl' _ DurGen to,a/a_hl_gT,,Dr _ _4eaclGuarter _ Ser,_ces, Curectorate for mforma_lorl ODerat_on' _ and ReoK)r%, 12'15 Je_fE, r_on