The papers in this special section examine important current topics on multimodal data fusion in the medical context. All clinical data, including genomic and proteomic, play a role in the diagnosis and in particular in the treatment planning and follow-up. This is true for all types of data analyses whether in classification, regression, retrieval, clustering, or other. The interaction between several types of information is not always well understood. Experienced clinicians automatically and even unconsciously add multiple sources of information into their decision process, but machine learning tools often concentrate on single information sources. This special issue presents five examples where several data sources are fused. The papers give several examples of fusion techniques and also the results obtained in quite different application scenarios.

In this paper the effect of the dimensionality of data sets on the exploitation of synergy among known nearest neighbor editing and condensing tools is analyzed using a synthetic data set. The synergy is exploited through a tandem mode effusion approach that combines the proximity graph (PG) based editing scheme and the minimal consistent set (MCS) condensing technique. These two methods were selected on the basis of prior experience to representatively evaluate the effect of the data dimensionality. The algorithm level fusion of PG editing and MCS condensing is experimentally shown to be a powerful implement across the range of data dimensionality.

Sensor fusion models have been characterized in the literature in a number of distinctly different ways: in terms of information levels at which the fusion is accomplished; the objectives of the fusion process, the application domain; the types of sensors employed, the sensor suite configuration and so on. The characterization most commonly encountered in the rapidly growing sensor fusion literature based on level of detail in the information is that of the now well known triplet: data level, feature level, and decision level. We consider here a generalized input-output (I/O) descriptor pair based characterization of the sensor fusion process that can be looked upon as a natural out growth of the trilevel characterization. The fusion system design philosophy expounded here is that an exhaustive exploitation of the sensor fusion potential should explore fusion under all of the different I/O-based fusion modes conceivable under such a characterization. Fusion system architectures designed to permit such exploitation offer the requisite flexibility for developing the most effective fusion system designs for a given application. A second facet of this exploitation is aimed at exploring the new concept of self-improving multisensor fusion system architectures wherein the central (fusion system) and focal (individual sensor subsystems) decision makers mutually enhance the other's performance by providing reinforced learning. A third facet is that of investigating fusion system architectures for environments wherein the different local decision makers may only be capable of narrower decisions that span only a subset of decision choices. The paper discusses these flexible fusion system architectures along with related issues and illustrates them with examples of their application to real-world scenarios.

This study presents an effective approach to the hitherto little addressed problem of feature assessment and selection for pattern recognition in imprecisely supervised environments. Unlike in classical supervised environments wherein the representative training samples have crisp class labels, here the samples have fuzzy memberships in several of the different pattern classes in the environment. The new methodology reported here is an outgrowth of a recently developed tool CORPS—Class Overlap Region Partitioning Scheme initially designed for operation in supervised environments and extended later for operation in imperfectly supervised environments. The emphasis here has been the development of a computationally efficient scheme capable of evaluating as rapidly as practical a large number of features individually as to their discrimination potential based on which a smaller subset may be selected, if so desired, for more detailed evaluation in different combinations by other tools.

A new approach is presented in this study for tackling the problem of high computational demands of nearest neighbor (NN) based decision systems. The approach, based on the concept of an optimal subset selection from a given training data set, derives a consistent subset which is aimed to be minimal in size. This minimal consistent subset (MCS) selection, in contrast to most of the other previous attempts of this nature, leads to an unique solution irrespective of the initial order of presentation of the data. Further, consistency property is assured at every iteration. Also, unlike under most prior approaches, the samples are selected here in the order of significance of their contribution for enabling the consistency property. This provides insight into the relative significance of the samples in the training set. Experimental results based on a number of independent training and test data sets are presented and discussed to illustrate the methodology and bring to focus its benefits. These results show that the nearest neighbor decision system performance suffers little degradation when the given large training set is replaced by its much smaller MCS in the operational phase of testing with an independent test set. A direct experimental comparison with a prior approach is also furnished to further strengthen the case for the new methodology.>

The concepts of fuzzy learning and learning in vicissitudinous environments are synergistically combined to develop a new fuzzy logic based approach to the problem of pattern recognition in variably supervised environments. The fuzziness imposed on the recognition process by the varying imperfectness in the supervision is modeled with fuzzy membership functions. The learning under the environment is simultaneously accompanied by learning about the vicissitudinous nature of the environment and the fuzzy membership functions.>

Efficient decision fusion strategies for deriving optimal decisions in multisensor target recognition/tracking environments are analyzed. The problem is viewed as one of parallel fusion of the individual independent decisions with a recursive structure that permits enhancement of the reliability of the fused decision. Instead of a voting-based fusion approach, this recursive structure can be combined with any one of the other optimal fusion techniques as well. An analysis based on a simple voting or consistency-measure-dependent fusion scheme is presented to illustrate the recursive model and its benefits.>

The problem of recognition in nonparametric environments under imperfect supervision is not amenable to solution through classical statistical approaches based on identification of finite mixtures, which require an a priori knowledge of the probabilistic descriptions of the classes. Accordingly, the problem is viewed in this study as one of optimal linear/nonlinear partitioning of the imperfectly labeled training sample set. This optimal partitioning is accomplished by defining an appropriate optimality criterion, which takes into account the imperfectness of supervision, and solving the resultant optimization problem through the Improved Flexible Polyhedron Method (IFPM). Possible alternatives to compensate for the inherent bias in this criterion towards equipopulation clusters are developed and evaluated using an illustrative example. Details of the methodology involved in implementing the approach are presented. Results of simulation experiments, which confirm the validity and effectiveness of this new technique in accomplishing optimal, linear/nonlinear discriminant learning in imperfectly supervised, nonparametric environments, are included.

The scope of the classical k-NN classification techniques is enlarged under this study to cover partially exposed environments. The modified classification system structure required for successful operation in environments, wherein all the inherent pattern classes are not exposed to the system prior to deployment, is developed and illustrated with the aid of a specific classification rule-the neighborhood census rule (NCR). Admittedly, alternative rules can be visualized to fit this modified structure. However, this study concentrates on the use of NCR to bring out the underlying philosophy and develops optimum thresholds for admittance of unknown samples into the set of presently known classes. These thresholds are learned from the available training samples of these classes. This learning represents a new dimensionality of the learning system structure in that estimates of the domains of the known classes are developed in addition to learning of the discrimination among these classes. This facilitates identification of samples belonging to the classes previously unexposed to the recognition system. Experimental results are also presented in support of the proposed concepts and methodology for operation in partially exposed environments.

Generalized nonlinear dynamic models of bistate ecological processes and the feasibility of obtaining closed-form analytical solutions to their governing differential equations represent the topic of this study. The model considered here permits arbitrary nonlinearities within certain very broad limitations, and consequently the study can be of value in modeling and analysis of bistate processes with complex inherent non-linearities in their behavior. The methodology is illustrated by considering as an example a class of these processes wherein the populations of both the species are in the same mode of growth or decay at any given instant. Exact closed-form solutions for the dynamics of a particular process belonging to this class are derived analytically. The resulting responses are discussed to bring out the behavior of the process in all its complexity and thereby demonstrate the value of this technique as an analytical tool in the study of a wide spectrum of nonlinear, bistate, socioecological processes. Also, it is shown that the scope of this study can be enlarged further to cover classes of processes with time-dependent properties as well.

Here the twin problems of feature selection and learning are tackled simultaneously to obtain a unified approach to the problem of pattern recognition in an unsupervised environment. This is achieved by combining a feature selection scheme based on the stochastic learning automata model with an unsupervised learning scheme such as learning with a probabilistic teacher. Test implementation of this scheme using the remotely sensed agricultural data of the Purdue laboratory for agricultural remote sensing (LARS) in a simulated unsupervised mode, has brought out the efficacy of this integrated system of feature selection and learning.