Stephen Shervais

Intelligent supply chain management using adaptive critic learning

A set of neural networks is employed to develop control policies that are better than fixed, theoretically optimal policies, when applied to a combined physical inventory and distribution system in a nonstationary demand environment. Specifically, we show that model-based adaptive critic approximate dynamic programming techniques can be used with systems characterized by discrete valued states and controls. The control policies embodied by the trained neural networks outperformed the best, fixed policies (found by either linear programming or genetic algorithms) in a high-penalty cost environment with time-varying demand.

Reconstructability Analysis as a Tool for Identifying Gene-Gene Interactions in Studies of Human Diseases

There are a number of common human diseases for which the genetic component may include an epistatic interaction of multiple genes. Detecting these interactions with standard statistical tools is difficult because there may be an interaction effect, but minimal or no main effect. Reconstructability analysis (RA) uses Shannons information theory to detect relationships between variables in categorical datasets. We applied RA to simulated data for five different models of gene-gene interaction, and find that even with heritability levels as low as 0.008, and with the inclusion of 50 non-associated genes in the dataset, we can identify the interacting gene pairs with an accuracy of ?80%. We applied RA to a real dataset of type 2 non-insulin-dependent diabetes (NIDDM) cases and controls, and closely approximated the results of more conventional single SNP disease association studies. In addition, we replicated prior evidence for epistatic interactions between SNPs on chromosomes 2 and 15.

Reconstructability analysis detection of optimal gene order in genetic algorithms

The building block hypothesis implies that genetic algorithm efficiency will be improved if sets of genes that improve fitness through epistatic interaction are near to one another on the chromosome. We demonstrate this effect with a simple problem, and show that information‐theoretic reconstructability analysis can be used to decide on optimal gene ordering.

Reconstructability analysis of genetic loci associated with Alzheimer disease

Reconstructability Analysis (RA) is an information- and graph-theory-based method which has been successfully used in previous genomic studies. Here we apply it to genetic (14 SNPs) and non-genetic (Education, Age, Gender) data on Alzheimer disease in a well-characterized Case/Control sample of 424 individuals. We confirm the importance of APOE as a predictor of the disease, and identify one non-genetic factor, Education, and two SNPs, one in BINI and the other in SORCS1, as likely disease predictors. SORCS1 appears to be a common risk factor for people with or without APOE. We also identify a possible interaction effect between Education and BINI. Methodologically, we introduce and use to advantage some more powerful features of RA not used in prior genomic studies.

Heuristic optimization as a V&V tool for software process simulation models

This work illustrates the use of heuristic algorithms to improve the verification and validation of software process simulation models. To use this approach, an optimization problem is formulated to guide a heuristic search algorithm that will attempt to locate particular combinations of parameter values that yield surprising results. These surprising results often help the modeler to identify flaws in the model logic that would otherwise remain undetected. The general concepts are discussed and a simple example is provided. Copyright

Ordering Genetic Algorithm Genomes with Reconstructability Analysis

The building block hypothesis implies that genetic algorithm (GA) effectiveness is influenced by the relative location of epistatic genes on the chromosome. We demonstrate this effect in four experiments, where chromosomes with adjacent epistatic genes provide improved results over chromosomes with separated epistatic genes. We also show that information-theoretic reconstructability analysis can be used to decide on optimal gene ordering.

Improving theoretically-optimal and quasi-optimal inventory and transportation policies using adaptive critic based approximate dynamic programming

We demonstrate the possibility of improving on theoretically-optimal fixed policies for control of physical inventory systems in a nonstationary fitness terrain, based on the combined application of evolutionary search and adaptive critic terrain following. We show that adaptive critic based approximate dynamic programming techniques based on plant-controller Jacobians can be used with systems characterized by discrete valued states and controls. Improvements over the best fixed policies (found using either an LP model or a genetic algorithm) in a high-penalty environment, average 83% under conditions both of stationary and nonstationary demand using real world data.

Reconstructability analysis as a tool for identifying gene-gene interactions in studies of human diseases

There are a number of human diseases that are caused by the epistatic interaction of multiple genes. Detecting these interactions with standard statistical tools is difficult, because there may be an interaction effect, but minimal or no main effect. Reconstructability analysis uses Shannons information theory to detect relationships between variables in categorical datasets. We apply reconstructability analysis to data generated by five different models of gene-gene interaction, with heritability levels from 0.053 to 0.008, using 200 controls and 200 cases. We find that even with heritability levels as low as 0.008, and with the inclusion of 50 non-associated genes in the data-set, we can identify the interacting gene pairs with an accuracy of 80% or better.

Using reconstructability analysis to select input variables for artificial neural networks

We demonstrate the use of reconstructability analysis to reduce the number of input variables for a neural network. Using the heart disease dataset we reduce the number of independent variables from 13 to two, while providing results that are statistically indistinguishable from those of NNs using the full variable set. We also demonstrate that rule lookup tables obtained directly from the data for the RA models are almost as effective as NNs trained on model variables.

Satisficing vs exploring when learning a constrained environment

Satisficing is an efficient strategy for applying existing knowledge in a complex, constrained, environment. We present a set of agent-based simulations that demonstrate a higher payoff for satisficing strategies than for exploring strategies when using approximate dynamic programming methods for learning complex environments. In our constrained learning environment, satisficing agents outperformed exploring agent by approximately six percent, in terms of the number of tasks completed.