Real world pattern recognition, model-based clustering, and many related problems are difficult because patterns vary and they occur in context, surrounded or partially occluded by other patterns. In simple cases, virtually any pattern recognition algorithm or neural network will give the same solution. In complex cases, all well known algorithms and neural network lead to combinatorial explosion (astronomical and worse numbers of training samples, or numbers of matches between pattern parametric models and data, etc.) Dynamic logic and neural modeling fields overcome this difficulty, reduce complexity to manageable size, and therefore are scalable to complex problems. Results usually reach information-theoretic performance limits (Cramer-Rao Bounds). Patterns can be found in noise and clutter, when signal-to-noise ratio is hundreds of times lower than when using other algorithms. Some recent publications:

Autonomous Intelligent Sensor Networks

See Program Management, Semantic Web

Detection of Patterns in Images and Signals

R. Deming, S. Higbee, D. Dwyer, M. Welser, L. Perlovsky, and P. Pellegrini, Robust detection and spectrum estimation of multiple sources from rotating-prism spectrometer images, Proceedings of SPIE Remote Sensing Europe, Stockholm, Sweden, 2006.

Tracking and Detection below noise and clutter

Deming, R.W. and Perlovsky, L.I. (2006). GMTI Tracking Improvement by 20dB, 52nd Annual Meeting of the MSS Tri-Service Radar Symposium, 19-23 June, 2006, Lincoln Laboratory, Lexington, MA.

Perlovsky, L.I. (2007). Neural Networks for Improved Tracking, IEEE Trans. Neural Netowrks. In print

Fusion of Signals from multiple platforms

Deming, R.W. and Perlovsky, L.I. (2007). Sensor Fusion for Swarms of Unmanned Aerial Vehicles using Fuzzy Dynamic Logic, Information Fusion, in print.

Radar Imaging through Walls

Kozma R., Linnehan R., Schindler J., and Perlovsky, L.I. (2007). Radar Imaging through Walls. IJCNN, Orlando, FL.

Search Engines with learning language ability

Perlovsky, L.I. (2006). Symbols: Integrated Cognition and Language. Chapter in Semiotics and Intelligent Systems Development. Eds. A. Loula, R. Gudwin, J. Queiroz. Idea Group, Hershey, PA, pp.121-151.

Integration of language with cognition and behavior

Tikhanoff, V., Fontanari, J.F., Cangelosi A., Perlovsky, L.P. (2006). Language and Cognition Integration through Modeling Field Theory: Category Formation for Symbol Grounding. International Conference on Artificial Neural Networks (ICANN’06), Athens, Greece .

Fontanari, J.F. and Perlovsky, L.I. (2007). Evolving Compositionality in Evolutionary Language Games. IEEE Transactions on Evolutionary Computations. In print.

Perlovsky, L.I. (2007). Fusion of Sensor and Language Signals at Pre-Conceptual Level, Information Sciences. In print.

Cramer-Rao Bounds for optimal system design

Cramer-Rao Bounds (CRB) makes possible to estimate limits on system performance as a function of its parameters, independent of any specific algorithm. We use CRB to optimize system design.

Robert Linnehan, David Brady, John Schindler, Leonid Perlovsky, Muralidhar Rangaswamy. (2007). On the Design of SAR Apertures Using the Cram´er-Rao Bound. IEEE Transactions on Aerospace and Electronic Systems. In print, Feb., 2007.

Financial Predictions  

I am affiliated with Ascent Capital Management, a company that uses dynamic logic and neural modeling fields for financial market prediction. Below is a recent weekly investment advise letter along with predicted portfolio performances.

PAST PERFORMANCE

on 2004 Dec 31, average annual cumulative performance, of the recommended SP500-portf. vs S&P500 market portfolio

5 year average

18.91%

-5.17%

3 year average 

3.47%

1.85%

1 year average 10.19% 8.68%

total cumulative portfolio gain since inception on Aug 1998 to Dec 2005
73.6%             (vs. S&P500  8.5%)

Annual results for 2005
recommended SP500-portf. = 0.5% gain (vs.market S&P500 = 3.0%  gain)
recommended NQ100-portf. = -2.5% loss (vs.market NQ 100 = 1.5% gain)
Maximum draw down: recommended 3.3%, 7.6%; (vs. market 5.7%, 13.1%)
Average exposure: recommended-portf.= 48%, 53%, vs. mkt-portf.= 100%
The 2005 annual performance of the recommended portfolios is a little worse than the market, with significantly lower risk

Annual results for 2006
recommended SP500-portf. = 12.3% gain (vs.market S&P500 = 13.6% gain)
recommended NQ100-portf. = 10.5% gain (vs.market NQ 100 =  6.8% gain)
Risk measures:
gain/st.dev. = 3.6, 1.9 (vs. mkts 1.3,  0.4)
Maximum draw down: recommended 1.3%, 3.1%;(vs. market  7.2%, 17.9%)
Average exposure: recommended-portf.= 26%, 24%; (vs. mkt-portf.= 100%)
(Assumption: part of portfolio not invested in futures is invested in 90-day T-Bills at 5% rate)

CURRENT PREDICTION
2007/02/09

*********************************************************************

*********************************************************************

!!! Model recommendations for close of 2007/02/16
S&P500 confidence level, PBS = -50%   

Short, average certainty

NASDAQ confidence level, PBS = -20%   
Short, low certainty

*********************************************************************

low abs(PBS) indicates (1) low certainty of our predictions and
(2) high volatility vs. directionality of the market.
Long or short positions proportional to PBS are advised.

Credits to Val Petrov and Ilya Perlovsky for contributions to model development and market analysis.