International Journal on Artificial Intelligence Tools

In this paper we propose a novel Support Vector Machines-based architecture for medical diagnosis using multi-class gene expression data. It consists of a pre-processing unit and N-1 sequentially ordered blocks capable of classifying N classes in a cascading manner. Each block embodies both a gene selection and a classification module. It offers the flexibility of constructing block-specific gene expression spaces and hypersurfaces for the discrimination of the different classes. The proposed architecture was applied for medical diagnostic tasks including prostate and lung cancer diagnosis. Its performance was evaluated by using a leave-one-out cross validation approach which avoids the bias introduced by the gene selection process. The results show that it provides high accuracy which in most cases exceeds the accuracy achieved by the popular one-vs-one and one-vs-all SVM combination schemes and Nearest-Neighbor classifiers. The cascading SVMs can be successfully applied as a medical diagnostic tool.

Exact determination of a gene network is required to discover the higher-order structures of an organism and to interpret its behavior. Most research work in learning gene networks either assumes that there is no time delay in gene expression or that there is a constant time delay. This paper shows how Bayesian Networks can be applied to represent multi-time delay relationships as well as directed loops. The intractability of the network learning algorithm is handled by using an improved mutual information criterion. Also, a new structure learning algorithm, "Learning By Modification", is proposed to learn the sparse structure of a gene network. The experimental results on synthetic data and real data show that our method is more accurate in determining the gene structure as compared to the traditional methods. Even transcriptional loops spanning over the whole cell can be detected by our algorithm.

We propose a self–adaptive probabilistic neural network model, which incorporates optimization algorithms to determine its spread parameters. The performance of the proposed model is investigated on two protein localization problems, as well as on two medical diagnostic tasks. Experimental results are compared with that of feedforward neural networks and support vector machines. Different sampling techniques are used and statistical tests are conducted to calculate the statistical significance of the results.

Based on the hypothesis that the sound of the infant cry contains information on the infant's health status, research has been done on how to improve classification of neonate crying sounds into categories called 'normal' and 'abnormal' - the latter referring to some hypoxia-related disorder. Research in this field is hindered by lack of test cases and limited understanding of feature relevance. The research described here combines various ways of dealing with the small data set problem. First, feature pre-selection is done using sequential backwards elimination of possible combinations where the performance of the set of features is tested by a Probabilistic Neural Network which has the advantage of fast learning. Using these features a neural network committee, consisting of Radial Basis Function Neural Networks, was trained on the data, using bootstrapping. This construction yields a multi-classifier system with an overall classification performance of 85% on the so-called "All Cry Units" (ACU) data set, an increase of 34% with respect to the a priori probability of 51%.

Several leave-1-out experiments for Linear Discriminant Analysis (LDA), Support Vector Machines (SVM) and Neural Networks (NN) have been conducted in order to compare the performance of the multi-classifier system.

Since the fetus is not available for direct observations, only indirect information can guide the obstetrician in charge. Electronic Fetal Monitoring (EFM) is widely used for assessing fetal well being. EFM involves detection of the Fetal Heart Rate (FHR) signal and the Uterine Activity (UA) signal. The most serious fetal incident is the hypoxic injury leading to cerebral palsy or even death, which is a condition that must be predicted and avoided. This research work proposes a new integrated method for feature extraction and classification of the FHR signal able to associate FHR with umbilical artery pH values at delivery. The proposed method introduces the use of the Discrete Wavelet Transform (DWT) to extract time-scale dependent features of the FHR signal and the use of Support Vector Machines (SVMs) for the categorization. The proposed methodology is tested on a data set of intrapartum recordings were the FHR categories are associated with umbilical artery pH values, This proposed approach achieved high overall classification performance proving its merits.

Significant advances have occurred in heuristic search for planning in the last eleven years. Many of these planners use A*-style search. We report on five sound and complete domain-independent forward state-space STRIPS planners in this paper. The planners are AWA* (Adjusted Weighted A*), MAWA* (Modified AWA*), AWA*-AC (AWA* with action conflict-based adjustment), AWA*-PD (AWA* with deleted preconditions-based adjustment), and AWA*-AC-LE (AWA*-AC with lazy evaluation). AWA* is the first planner to use node-dependent weighting in A*. MAWA*, AWA*-AC, AWA*-PD, and AWA*-AC-LE use conditional two-phase heuristic evaluation. MAWA* applies node-dependent weighting to a subset of the nodes in the fringe, after the two-phase evaluation. One novel idea in AWA*-AC-LE is lazy heuristic evaluation which does not construct relaxed plans to compute heuristic values for all nodes. We report on an empirical comparison of AWA*, MAWA*, AWA*-AC, AWA*-PD, and AWA*-AC-LE with classical planners AltAlt, FF, HSP-2 and STAN 4. Our variants of A* outperform these planners on several problems. The empirical evaluation shows that heuristic search planning is significantly benefitted by node-dependent weighting, conditional two-phase heuristic evaluation and lazy evaluation. We report on the insights about inferior performance of our planners in some domains using the notion of waiting time. We discuss many other variants of A*, state-space planners and directions for future work.

Semantic web mining is getting more attention in intelligent web applications. Many web sites, especially those dynamically generate HTML pages to display the results of user queries, present information in the form of lists or tables. It is very useful to extract concept instances from these tables for many web applications such as intelligent agent systems for on-line product recommendations. This paper describes a technique for extracting data from tables in two steps, namely table detection and table analysis. The table detection step identifies the existence of a table and extracts its contents, and the table analysis step discovers the semantic meanings embedded in the table and associates them with the concepts described in the domain ontology that are used for semantic annotation on these tables. Our algorithm has been tested based on real-life web documents and the experimental results are encouraging.

Combinatorial auctions allow bidders to bid for items leading to more efficient allocations, but determining winners in auctions is -complete. In this work, a simple yet effective Lagrangian relaxation based heuristic algorithm is presented. Extensive computational experiments using standard benchmark data (CATS) as well as newly generated more realistic test sets were conducted which showed the heuristic was able to provide optimal solutions for most test cases and is within 1% from the optimums for the rest within very short times. Experiements comparing CPLEX 8.0, the fastest current algorithm, showed the heuristic was able to provide equally godd or better solutions often requring less than 1% of the time required by CPLEX 8.0.