Computational Intelligence, Volume 25, Number 3, 2009 A CBR-BASED, CLOSED-LOOP ARCHITECTURE FOR TEMPORAL ABSTRACTIONS CONFIGURATION STEFANIA MONTANI, 1 ALESSIO BOTTRIGHI, 1 GIORGIO LEONARDI, 2 AND LUIGI PORTINALE 1 1 Universit` a del Piemonte Orientale, Alessandria, Italy 2 Universit` a di Pavia, Pavia, Italy In the hemodialysis domain, we are implementing a case-based, closed-loop architecture aimed at configuring temporal abstractions (TA), which will be applied to time series data. The advantage of a case-based approach is the one of “quickly” obtaining a suitable TA parameter configuration, simply by looking at the most similar already configured case, where configured cases are indexed by means of contextual information. The retrieved configuration, together with the time series data, is then used as an input to a TA processing module, able to provide a set of qualitative states, trends, and significant combinations of both as an output. TA processing results can finally be evaluated, possibly leading to a (human-supervized) reorganization/revision of the case base content, to ameliorate future TA configuration sessions—thus closing the loop. The work is being integrated with RHENE, a system for case-based retrieval in hemodialysis, able to work both on raw time series data and on preprocessed (by means of TA) ones. Key words: case-based reasoning, temporal abstractions, parameter setting. 1. INTRODUCTION Parameter configuration is a critical issue in many artificial intelligence (AI) processes, especially when they are applied to complex domains like medical ones. The temporal abstractions (TA) (Shahar 1997; Bellazzi, Larizza, and Riva 1998) method- ology, in particular, is an AI process that requires a nontrivial configuration phase, because parameter values highly influence TA output. TA are resorted to map large amounts of temporal information, such as the ones embed- ded in a time series, to a compact representation, able not only to summarize the original data themselves, but also to abstract and highlight meaningful behaviors in them. The basic principle of TA methods is to move from a point-based to an interval-based representation of the data, where: (i) the input points are the elements of the discretized time series; (ii) the output intervals (also called episodes henceforth) aggregate adjacent points sharing a common behavior, persistent over time. More precisely, the method described above should be referred to as basic TA (Bellazzi et al. 1998). Basic abstractions can be further subdi- vided into state TA and trend TA. State TA are used to extract episodes associated with qualitative levels of the monitored feature, e.g., low, normal, high values; trend TA are ex- ploited to detect specific patterns, such as increase, decrease or stability, in the time series. Complex TA (Bellazzi et al. 1998), on the other hand, aggregate two series of episodes into a set of higher level episodes (i.e., they abstract output intervals over precalculated input intervals). In particular, they are used to search for specific temporal relationships between episodes that can be generated from a basic abstraction or from other complex abstrac- tions. The relation between time intervals can be any of the temporal relations defined by Allen (Allen 1984). This kind of TA can be exploited to extract patterns that depend on the course of several features, or to detect patterns of complex shapes (e.g., a peak) in a single feature. TA configuration usually demands for domain knowledge, which could be unavailable, or whose elicitation, exploitation and maintenance could be too time-consuming in practice. Address correspondence to Stefania Montani, Dipartimento di Informatica, Universit` a del Piemonte Orientale, Via T. Michel 11 I-15121 Alessandria, Italy; e-mail: stefania.montani@unipmn.it C  2009 The Authors. Journal Compilation C  2009 Wiley Periodicals, Inc.