|
V. William Porto |
Raw data can be filtered, aggregated, transformed, visualized, and correlated. However, data by itself is of no real value. For data to be of value it must first be interpreted. Interpretation/analysis is a necessary but insufficient condition for conveying information. The interpretation must impact the behavior of the end user for the data to be truly useful. This forms the underlying basis of all information processing systems: utilize data collected from a variety of sources to piece together a view of the world which enables an end user to make enlightened decisions with respect to this environment.
Information processing (IP) systems take various forms, many of which combine technology from several scientific fields. Within the past decade we have seen IP systems incorporate the latest technological advances in artificial neural networks (ANNs), fuzzy logic, and genetic algorithms/evolutionary computation. Artificial neural networks are perhaps the most well publicized technology of this group, with application areas ranging from control systems, pattern recognition/ classification, and modeling systems to fault diagnosis and adaptive filtering. The ocean sciences have benefited significantly from these technologies as they have provided new solutions to difficult and previously (computationally) intractable problems. This article highlights neural network and information processing efforts in oceans technology over the past decade, and presents some predictions about future developments in this dynamically changing arena.
Artificial neural networks are mathematical models originally designed to mimic aspects of how we believe the brain works. Neural networks are parallel processing structures consisting of non-linear processing elements interconnected by fixed or variable weights. Computation nodes typically sum N weighted inputs and pass the result through a nonlinear function. Data flows through this series of non-linear transformations (e.g., sigmoid), in one or more stages in an assigned topology. Neural network topologies can be structured to generate arbitrarily complex decision regions for stimulus-response pairs, hence they are perfectly suited for mapping input-output relationships. They provide the ability to construct complex, adaptive nonlinear filters that in many cases outperform linear discriminant functions. Linear discriminant mapping functions were previously designed and used because of their mathematical tractability.
One of the significant advantages of neural networks lies in their parallel distributed processing structures. Rather than performing a programmed set of instructions sequentially as in a traditional Von Neumann type computer, neural network nodal functions can be evaluated simultaneously, thereby gaining enormous increases in processing speed. Although software representations of actual neural networks are typically used, special purpose neural-processor boards have also recently become available.
The benefits of using artificial neural networks to solve real-world problems extend beyond the high computational rates provided by their massive parallelism. Neural network classifiers are non-parametric and make weak or no assumptions about the probabilistic properties of the underlying distributions of the data. Traditional detector/ classifiers require assumptions concerning the data distributions and may undergo a degradation in performance when these assumptions are violated. Thus, neural network classifiers are more robust when distributions of the data are generated by nonlinear processes and are strongly non-Gaussian.
Neural network paradigms can be divided into two basic categories: supervised learning and unsupervised learning structures. In a supervised learning paradigm, the input data is associated with some output criterion in a one-to- one mapping, with this mapping known a priori. This mapping is learned by the network in the training phase and thus similar inputs are associated with the various desired output classes. Input vectors may consist of frequency components, pixel values, transform coefficients, or any other features considered important. Complex decision regions are typically formed using hyperplane decision boundaries or kernel-function nodes that form overlapping receptive fields.
Unsupervised learning paradigms form decision regions based on whether or not an input exemplar is sufficiently similar to the exemplars previously learned. No training supervision is needed, eliminating the need for a priori input-output pair associations. Various metrics can be utilized for determination of goodness of fit. If the similarity measure exceeds a threshold, a new class of exemplars is created and this process is then repeated for each new input exemplar. These network topologies are often used for clustering data and are quite useful where knowledge of unknown associations existing between input exemplars is needed.
Finally, neural networks must be trained prior to use. Training a neural network effectively synthesizes a set of rules from a body of training exemplars. During the training phase, the neural network encodes the necessary transformation, mapping a desired set of input features to specific output features. Applicable training methods are a function of the characteristics of the neural topology and nodal functions.
There are a wide variety of areas in which artificial neural networks have been applied to problems in the ocean sciences. These include pattern recognition, optimal control, adaptive filtering, inversion, target tracking, and general purpose modeling, among many other imaginative applications.
Modern control theory has scored impressive gains in the comparatively few years of its existence. Improvements have been made in conventional optimal control, as well as many forms of adaptive control, pole-placement using state-variable feedback, control of multi-input/multi-output systems by entire eigenstructure assignment, and Kalman filtering. As powerful as the theory has become, it still has its limitations. These are often felt when dealing with time varying systems, such as those encountered in navigating an autonomous underwater vehicle. Nonlinear environments can pose significant difficulties. Systems designed by optimal control theory are not always robust with respect to parameter variation or changes in the form of the input relative to that of the original design. To be truly useful, a navigation controller must learn and encode control changes through a variety of environments while also learning new state inputs and required corrections. Conventional controllers and expert systems often have difficulty in dealing with these nonlinear environments.
Neural networks offer a distinctly different approach to optimal control. Because they are robust, they are particularly well suited to controlling systems in dynamic non-linear environments. Neural networks also offer the ability to design nonlinear controllers without requiring complex system models. Several testbed environments have been developed to assist in developing robust controllers [1][5]. Tracking errors are typically used in a feedback loop wherein the ANN learns the appropriate responses to the input stimuli. In one of these systems, a neural-controller environment (SIGNAL) has been created to specifically assist in development of a variety of marine vessel control loops [1]. This testbed can be used to design a number of control systems ranging from navigation to ballast stabilization systems. It is currently being used to design autonomous subsea robot controllers since it can utilize real-time probes and learn in-situ.
Other control applications include depth controllers for UUVs [2], current prediction in shallow water environments [3], and adaptive heave compensation [4]. More recently, an ANN has been used to synthesize a robust velocity control system for a Remotely Operated Vehicle (ROV). Researchers Pollini and Nasuti [6] incorporated a feedback linearization controller (FBLC) in conjunction with a neural network. Mixed pattern batch learning was used together with linearized output nodes to greatly increase the convergence rates and facilitate on-line learning. Though the use of linearized activation functions makes this approach similar to standard adaptive controllers, the use of multiple, redundant neurons in the topology makes the system performance much more robust.
One of the largest and most successful application areas involves the use of ANNs for pattern recognition, classification and detection. The goal of pattern recognition and classification is that of assigning each separate input pattern to one of a finite number of output classes. Input elements represent measurements of selected features that are used to distinguish between the various output classes. These input patterns form a multi-dimensional feature space, with the job of the classifier being to partition this space into decision regions indicative of the class to which each input pattern belongs. Quite a variety of marine classification/detection problems have been attacked and solved through the use of ANNs.
One early study incorporated an ANN to obtain a low false alarm rate in a passive sonar transient detection and classification system. Researchers at GTE Government Systems [7] used a two-stage approach wherein a transient processing chain is used to detect and classify short duration transients together with a tonal processing chain to handle longer duration transient signals. Power spectral estimates with two separate pre-processing stages generated the input features to the ANNs. A false alarm rate of 5 x 10-6 was achieved with a standard transient data test set. The ANNs were trained in the presence of additive noise. Results show that all of the incorrect classifications (in this test set) were due to missed detections. It is noteworthy to mention that the learning mechanism can be adjusted to minimize type 1 or type 2 errors (false positive, false negative) as required by the application.
ANNs have been used to interpolate seafloor sediments where only sparse measurements are available. Caiti and Parisini [8] were able to design a neural network that was able to generate a continuous mapping of various sediment properties as a function of the three-dimensional position of the sediment measurements. Radial-basis function (RBF) node topologies were used instead of the often-used multi-layer perceptron (MLP) structures. RBF structures are well suited to this application. RBFs form smooth mappings and do not rely on the specifics of the measurement system or physical properties, hence measurements from a variety of sensors can even be utilized simultaneously. Results of this study show the network was able to generate predictions with very low mean square error (with their data sets).
More recently, neural networks were used to classify marine sediments using measurements from a sub-bottom profile model [12]. The goal of this research was to accurately classify the topmost layer in a heterogeneous sediment profile. A two stage feature extraction method was combined with a RBF network classifier. Raw acoustic signals are poorly suited for use as input features to a classifier because useful information is buried within their structure. Feature extractors were designed using wavelet packet analysis, eigen analysis, and a non-linear component based upon distances between desired class prototypes and the input vectors. All of these features were passed through a discriminant factor analysis tool to obtain the final input vector for the ANN. Results using computer simulated data are quite promising. Statistical results show that relatively high classification rates can be achieved via this approach, potentially eliminating the need for costly and tedious direct sediment sampling.
Other ocean related classification/ detection problems tackled with neural networks include automatic object recognition systems using synthetic pattern imagery [9], pollutant induced tissue changes in fish livers [10], plankton recognition [10], and seafloor imagery [11]. Multi-layer perceptron topo- logies are often used in these studies due to their ease of training.
Buried object detection using ANN techniques has also studied recently. Researchers in Italy exploited object resonances induced by an active sonar ping [13]. A classify-before-detect strategy was employed as it takes advantage of the microstructure in acoustic signatures. Both Maximum Likelihood Estimator (MLE) and Multivariate Gaussian Classifier (MVG) approaches were studied with time-frequency features used as inputs to the ANN. Efforts focused on training an ANN on the low frequency components of the parametric sonar signal. Instead of utilizing a traditional Euclidean distance metric, a Mahalanobis distance metric and a Bayesian distance metric were used in the MVG and MLE classifiers, respectively. Analysis of ROC curves illustrate the capability of this classify-before-detect technique to handle signals with low signal-to-reverberation ratios.
Neural networks have also been quite fruitful in ocean applications where optimal adaptive filters must be designed. In a paper by El-Hawary and Li [14], an ANN was designed and used to filter and classify a series of sinusoidal signals at different SNR levels (ranging from 0.1dB to 100dB). A feed-forward two hidden layer perceptron architecture was trained, and results compared to a standard linear prediction filter (LP). Their research showed that they were able to obtain results that are equal to or better than the LPC in all of the test cases. Importantly, they note that it is only necessary to train the ANN once with a clean signal whereas the LP filter must estimate the parameters for each set of signalS for a given order p. This is a prime example of the advantages of the inherent non-linear mapping properties of the ANN when compared to standard statistical (i.e., linear) techniques.
A related approach to designing non-linear filters was shown by Gomes and Barroso [15]. Radial basis function neural networks were used for blind adaptive equalization in a set of acoustic receivers. Blind equalization attempts to reduce inter-symbol interference and channel fading. There are many standard approaches to this problem, but each makes performance limiting assumptions regarding the underlying noise and transmission distortions. RBF networks were used in an unsupervised learning mode to cluster the input data. The performance of the RBF networks to solve this problem was evaluated using both simulated and real world data sets. Results show that the network is tolerant of significant errors in the placement of cluster centers, but may be limited to problems of low to moderate dimension. They note that other ANN topologies (e.g., recurrent networks) may alleviate the need for a large number of nodes in the network (which often leads to brittle classifiers).
An interesting approach to inverting geophysical data incorporates the use of neural networks to learn topographic mappings. Ocean acoustic tomography involves estimating sound velocity profiles (and temperature, salinity, etc.) by inversion of travel times of sound from source(s) to receiver(s). Neural networks can be used to generate the inverse mapping of the sound velocity profiles [16], [17], [18]. Gan [17] used such an approach to solve the inverse problem. In his study, a parabolic equation (PE) model was used in the forward problem. A feed-forward MLP network was presented as a method capable of recognizing (2-dimensional) sound velocity profiles.
In a similar approach, Stephan [16] utilized a multi-layer perceptron to learn the mappings between sound-speed environments and arrival time patterns. A ray-tracing model with varied degrees of additive noise was used to generate the input data. The MLP technique produced errors (min, max, and mean) which were lower than the traditional linear approach based upon canonical decomposition of sound velocity profiles in triangular kernels. As expected, results using the neural methodology were most impressive as the differences from linear profile regions increase. In addition, the ANN method was better at rejecting estimation errors in the presence of noise when compared to the linear method. This approach was extended later [18] and proven robust on a larger simulated data set from the North-Atlantic environment.
Research by Terre, Golenzer, and Solaiman employed both MLPs and Hopfield networks to track and identify ray acoustic arrivals [20]. A temporal series of arrivals was generated by tracking resolved peaks. Outputs of a ray-tracing model were processed together with peak time series to identify the arrival times. In addition to results which were favorable to traditional methods, significant improvements in processing speed were also realized. When compared to a Bayesian method, the neural system required less than ten percent of the time to process a Megapixel image.
Target tracking is another area in which ANNs have been quite useful. A group of researchers at NUSC developed a system which utilized five individual feed-forward networks trained to provide estimates of contact state variables given time series measurements [19]. Traditional target tracking methods (i.e., Extended Kalman filtering (EKF)) all make limiting assumptions, typically reducing to and solving the problem as a linearized dynamical system. In environments of poor observability, or when non-linear kinematic relationships between the observer and the contact exist (i.e., real-world systems), stability is a major concern. The series of five ANNs was able to learn the underlying dynamical system models throughout a range of observability conditions. When compared with traditional statistical methods, the ANNs demonstrated high levels of tracking accuracy using both azimuth bearing and range measurements, as well as in highly non-linear conical bearings-only problems.
Since neural networks can map arbitrary input-output associations, they are perfectly suited for modeling virtually any number of dynamical systems. One such problem is generating an interpolative/ extrapolative model of ocean data wherein the number of observations (i.e., salinity) is sparse or irregularly spaced. ANNs can be trained to learn optimal meshes of these data. Neural meshing of a geographical space relative to oceanographic data has been performed by Sarzeaud et al. using Kohonen self-organizing networks[21]. The nodes in the network were adapted by successively comparing the location of each datum and moving the closest node and its neighboring nodes toward this location.. Sarzeaud et al. were able to generate efficient neural meshes on a hydrological database of the North-East Atlantic, and point to the inherent capabilities of this method to generate meshes for a wide variety of data.
An auditory periphery neural model was designed by Dubrovsky and Rimskaya-Korsakova to detect and identify small objects via sonar [22]. Their neural model is based upon the auditory mechanisms in dolphins. Since the dynamic ranges of the neuron models were smaller than that of the signal sources, networks with different sensitivity thresholds combined with short- time adaptation mechanisms were used. Analysis of simulated echoes (of varying intensity) with three frequency channels was performed. Clustering of these multidimensional echo vectors (projected onto a plane) show excellent separation, with distinct boundary regions. This bionically inspired system was able to utilize highlights in weak echoes and was demonstrated to be quite effective in the classification of a variety of short impulsive echo types.
Chaotic modeling of radar backscattering from an ocean surface has also been performed via the use of neural networks. In a study by Haykin and Leung [23], radar backscatter (sea clutter) was modeled using radial basis function networks. This research used radar data to demonstrate the applicability of chaos theory for modeling sea clutter. A finite-dimensional deterministic model was developed using a RBF network as a predictor of the next value of the dynamical process. After the network was trained, it was shown to follow the sea clutter waveform closely. Importantly, through the use of their RBF models, the authors make the case that radar backscatter from the ocean surface may not be a stochastic process. Instead, they showed that a chaotic model learned by the RBF is not only plausible but also fits actual real-world data.
Neural networks have also been used to generate wind vectors from scatter- ometer data. Models developed by Mejia et al. utilized a multi-layer feed-forward network to learn the geophysical relationships between active microwave radar (transmitted and received via satellite) data and wind directions. Scattero- meters are used to calculate the normalized radar cross section of the ocean surface. Their neural model mapped the transfer function of an ERS-1 scattero- meter using over 30,000 wind vector pairs taken from North Atlantic ocean surveys. Subsequent testing produced results which outperform the previous statistical models, although in its current form the neural model’s dynamic range is smaller.
Neural networks have been used in a variety of ways to estimate image texture. In most of these studies, the network is used to make decisions on features extracted from imagery. Bourgeois and Walker [25], used a neural network to directly model image texture from sidescan sonar imagery. The task involved identification of homogeneous textures in an image. Standard image processing techniques (e.g., Fourier transforms, fractal modeling, etc.) often provide inadequate results. Many problems are due to the underlying assumptions made together with the fact that these techniques often discard data. Sub-image pixel data was used directly as inputs mapped to a set of desired image texture outputs. Results indicate the capability of multi-layer perceptrons to recognize input sequences even in the presence of significant amounts of random noise.
Segmentation of sidescan imagery has also been performed using a hybrid neural network approach. Research by the Deep Submergence Laboratory at Woods Hole Oceanographic Institution used a Markov random field model combined with a MLP to learn the distribution of observations from seafloor sidescan images[26]. Most seafloor acoustic imagery segmentation procedures rely on models that do not sufficiently take into account the spatial relationships of adjacent image regions. The authors’ neural network approach provides a non-parametric alternative to learn the probability contribution of each observation (pixels in an image set).
Another interesting application area involves using neural networks for fault diagnosis. In one study, a non-linear process model was designed to describe a set of fault types for a rudder control system [27]. A rudder-control system was monitored by an integral neural network fault-detection and identification model (FDI). The supervisory system is extended with an online learning, neural estimated fault model, and was continuously updated by differences between the actual system and the model state. A radial basis network generated accurate estimates of and feedback to the non-linear fault function, and was shown to potentially handle unanticipated faults (faults which were not previously modeled).
The aforementioned applications of neural networks to ocean science problems only hints at the potential for these techniques. Research in the past decade has proven the capability and utility of ANNs to solve a large number of problems. Previous techniques to solve many marine problems relied on linearization of the problem space (since linear models are easy to study) and mathematical models with limiting assumptions (i.e., Gaussian noise) because they were mathematically tractable. However, real-world environments are often highly non-linear. The ease in which neural networks can be designed and implemented makes them the obvious choice for numerous applications. Now that the groundwork has been laid and numerous ANN design toolsets have become available, we will see an expansion of their usage as they replace older, conventional methods. A longstanding concern that neural network solutions are difficult to analyze will be overshadowed by the fact that they provide proven, simple, and robust solutions in addition to performance advantages (e.g., computational speed and increased accuracy).
Predictions for the future include imbedded neural network systems which implement online in-situ learning instead of using weights fixed from previous training sessions. These will be important in navigation control, prediction of target dynamics, as well as for pattern recognition. Fault diagnosis and prediction is another upcoming area. Imagine an online neural system which can predict incipient failure of a hydraulic pump or propeller driveshaft with sufficient time to make necessary repairs. Models of marine life interactions, environmental pollution dynamics, and shallow-water surveillance are other promising areas ripe for the application of ANNs.
It is important to note that ANNs are only one tool in a multifaceted toolbox. In the future we will see ANNs integrated into hybrid systems. By combining the useful properties of neural networks together with evolutionary computation, fuzzy logic, and new visualization tools, ocean scientists will be able to solve many previously intractable problems. Future IP systems will depend on integration of several of these advanced tools to better solve problems. The inherent and unique capabilities of each tool will be exploited instead of relying on a singular technology to solve the entire problem.
[1] Cherruel, G., Y. Autret, and F. Dupont, (1994), "Using SIGNAL for developing Neural Control Systems," Proceedings of OCEANS ’94, pp. III - 116-120.
[2] Sutton, R., C. Johnson, and G.N. Roberts, (1994), "Depth Control of an Unmanned Underwater Vehicle Using Neural Networks," Proceedings of OCEANS ’94, pp. III - 121-125.
[3] Wüst, J.C., and G.J.H.L. van Noort, (1994), "Neural Network Current Prediction for Shipping Guidance," Proceedings of OCEANS ’94, pp. I - 58-63.
[4] Lainiotis, D.G., K.N. Plataniotis, D. Menon, and C.J. Charalampous, (1993), "Adaptive Heave Compensation Via Dynamic Neural Networks," Proceedings of OCEANS ’93, pp. I – 243-247.
[5] Guo, J., F.C. Chiu, and C. Wang, (1995), "Adaptive Control of an Autonomous Underwater Vehicle Testbed Using Neural Networks," Proceedings of OCEANS ’95, pp. 1033-1039.
[6] Pollini, L., M. Innocenti, and F. Nasuti, (1997), "Robust Feedback Linearization with Neural Network for Underwater Vehicle Control," Proceedings of OCEANS ’97, pp. 12-16.
[7] Casselman, F., D.F. Freeman, D.A. Kerrigan, S.E. Lane, N.H. Millström, and W.G. Nichols Jr., (1991), "A Neural Network-based Passive Sonar Detection and Classification Design with a Low False Alarm Rate," Proceedings of OCEANS ’91, pp. 1701-1705.
[8] Caiti, A. and T. Parisini, (1991), "Interpolation of Ocean Sediment Properties by Networks of Parallel Com- putational Units," Proceedings of OCEANS ’91, pp. 1695-1700.
[9] Menon, M., (1991), "Applications of Neural Networks to Automatic Object Recognition using both Imaging and Non-Imaging Data," Proceedings of OCEANS ’91, pp. 1375-1379.
[10] Ellis, R., R. Simpson, P.F. Culverhouse, T. Parisini, R. Williams, B. Reguera, and B. Moore, (1994), "Expert visual classification and neural networks: can general solutions be found," Proceedings of OCEANS ’94, pp. I – 330-334.
[11] Zerr, B., E. Maillard, and D. Gueriot, (1994), "Sea-Floor Classification by Neural Hybrid System," Proceedings of OCEANS ’94, pp. II – 239-243.
[12] Maroni, C.S., A. Quinquis, and E. Radoi, (1997), "A Methodology for Neural Network Based Classification of Marine Sediments using a Subbottom Profiler," Proceedings of OCEANS ’97, pp. 1370-1375.
[13] Granara, M., A. Pescetto, F. Repetto, G. Tacconi, and A. Trucco, (1998), "Statistical and Neural Techniques to Buried Object Detection and Classification," Proceedings of OCEANS ’98, pp. 1269-1273.
[14] El-Hawary, F., and J. Li, (1994), "Artificial Neural Network for Additive Noise Filtering Techniques," Proceedings of OCEANS ’94, pp. I – 323-329.
[15] Gomes, J., and V. Barroso, (1995), "Blind Equalization Using a Radial Basis Function Neural Network," Proceedings of OCEANS ’95, pp. 797-802.
[16] Stephan, Y., S. Thiria, and F. Badran, (1995), "Inverting Tomo- graphic Data with Neural Nets," Proceedings of OCEANS ’95, pp. 1501-1504.
[17] Gan, W.S., (1991), "Applications of Neural Networks to Ocean Acoustic Tomography," Proceedings of OCEANS ’91, pp. 1707-1712.
[18] Stephan, Y., (1996), "Neural Inversion of Gastom’90 Tomographic Data," Proceedings of OCEANS ’96, pp. 656-659.
[19] Ferkinhoff, D.J., C.T. Nguyen, S.E. Hammel, and K.F. Gong, (1993), "Performance Characterization of Artificial Neural Networks for Contact Tracking," Proceedings of OCEANS ’93, pp. II – 271-276.
[20] Terre, T., J. Golenzer, and B. Solaiman, (1995), "Tracking and Identification of Ray Acoustic Arrivals by Means of Neural Networks," Proceedings of OCEANS ’95, pp. 1454-1458.
[21] Sarzeaud, O., Y. Stéphan, F. LeCorte, and L. Kerléguer, (1994), "Neural Meshing of a Geographical Space in Regard to Oceanographic Data Location," Proceedings of OCEANS ’94, pp. I – 335-339.
[22] Dubrovsky, N.A., and L.K. Rimskaya-Korsakova, (1994), "A Simulation Network of First Order Auditory Neurons for Preprocessing of Acoustic Signals," Proceedings of OCEANS ’94, pp. II – 235-238.
[23] Haykin, S., and H. Leung, (1991), "Chaotic Modeling of Radar Backscatter from an Ocean Surface," Proceedings of OCEANS ’91, pp. 1713-1715.
[24] Mejia, C., S. Thiria, M. Crépon, and F. Badran, (1994), "A Neural Network Approach for Wind Retrieval from the ERS-1 Scatterometer Data," Proceedings of OCEANS ’94, pp. I – 76-80.
[25] Bourgeois, B., and C. Walker, (1991), "Sidescan Sonar Image Interpretation with Neural Networks," Proceedings of OCEANS ’91, pp. 1687-1693.
[26] Jiang, M., W.K. Stewart, and M. Marra, (1993), "Segmentation of Seafloor Sidescan Imagery Using Markov Random Fields and Neural Networks," Proceedings of OCEANS ’93, pp. III – 456-461.
[27] Vukic, Z., D. Pavlekovic, and H. Ozbolt, (1998), "Rudder Servo-System Fault Diagnosis Using Neural Network Fault Modeling," Proceedings of OCEANS ’98, pp. 538-543.
Mr. Porto has been involved in the OES for the past 10 years and currently serves as IEEE OES Technology Committee Chairman for Neural Networks and Information Processing. He did his undergraduate and graduate work (B.A and M.A) in mathematics and applied mathematics at the University of California at San Diego. He has worked in such areas as optimal sensor control, multi-hypothesis multi-sensor target tracking, sensor fusion, and transient signal processing. More recently he has concentrated his efforts in the field of computational intelligence (neural networks, fuzzy logic, and evolutionary computation). Mr. Porto has published over a number of papers in refereed journals and conferences. He is currently Vice President of Natural Selection, Inc. in La Jolla, California.
Mr. Porto has served on the Office of Naval Research advisory panel for neural networks, and is a charter officer of the Evolutionary Programming Society. He was on the organizing committees for the 1988 and 1989 IJCNN conferences, served as tutorials co-chair for OCEANS’ 95, and more recently was general chairman of the 1998 Conference on Evolutionary Programming. Mr. Porto is a member of the editorial board for the IOP/Oxford Press Handbooks for Neural Computation and Evolutionary Computation and is an associate editor for the IEEE Transactions on Evolutionary Computation.