Syracuse Evolutionary and Neural Systems Exploration (SENSE) research group


WHO?

Profs. Kishan Mehrotra and Chilukuri K. Mohan lead the group. Current doctoral students in the research group include Buthainah Al-kazemi (discrete particle swarm optimization), Kamala Anupindi (financial optimization using evolutionary algorithms), June Phansiri Athikomrungsarit (uncertainty manipulation in the context of multiple evidence sources that supply data intermittently), S. Kanat Bolazar (diversity in evolution strategies), Stu Card (genetic programming and wavelet networks), and David Walter (fuzzy classifier systems for diagnosis tasks).

Recent doctoral graduates, in reverse chronological order, include:

Research Interests:

We study evolutionary algorithms and artificial neural networks (see Publicationsfor details on specific papers). MIT Press has recently published an introductory textbook, Elements of Neural Networks, by Profs. Mehrotra, Mohan and Ranka.

Some of our recent results are in the development of new, general-purpose crossover operators.

Selective Crossover performs slight perturbations in parent individuals to compute resulting changes in parent fitness, and uses this information to generate new offspring. Experimental results confirm that this leads to rapid improvement in average and best fitness, when compared to 1PTX, 2PTX, and UX. For example code, please see SX.c for a deceptive problem
(Paper appeared in the 1998 Symposium on Applied Computing (SAC'98), February 27-March 1, 1998. More recent results are reported in IEEE-CEC, July 1999. Current work has applied this approach to optimization problems with significantly overlapping deceptive `building blocks'.)

Linkage Crossover applies a probabilistic inference methodology to crossover, with explicit representation of first order linkages between chromosomal components. For problems in which little/no linkage information is available a priori, we show that a Hebbian learning rule can be applied to learn appropriate linkages. The success of this approach is shown in 3 ways:
(i) performance (best fitness) is improved,
(ii) linkage adaptation process does converge, and
(iii) learnt linkage probabilities do correspond to expectation (for problems whose linkage structure is known).
(Paper appeared in the 1998 Symposium on Applied Computing (SAC'98), February 27-March 1, 1998, Atlanta (GA). More recent results are reported in GECCO, July 1999. Extensive experiments are discussed in a forthcoming paper in Evolutionary Computation.

In Adaptive Multi-Module Approximation Networks, we present a new adaptive algorithm for function approximation tasks. The learning algorithm successively introduces new (small) neural network modules, and associates each module with a set of reference vectors identifying its most appropriate region of applicability. This approach is especially useful in approximating discontinuous functions and function collages; sub-functions may also be smooth.


[Come back later to this page, I'll be adding more information soon.]

Mailing address: Prof. C.K.Mohan, 2-171 CST, Dept. of EECS, Syracuse University, Syracuse, NY 13244-4100, USA`

Send email to ckmohan@syr.edu or visit the web-page of Chilukuri K. Mohan