Thursday, September 17, 2015

WELCOME ...

to the Cognitive Science Modeling Lab (CSML) at the University of Evansville. CSML is a small lab devoted to agent-based modeling and, primarily, to developing virtual circuits that can perform rudimentary cognitive tasks. Our virtual circuits are also called by us "Dynamic Associative Networks" (DANs) to distinguish them from the more traditional Artificial Neural Networks (ANNs). DANs differ from ANNs in that they use no predetermined network structure; rather, we rather promiscuously add nodes in the network wherever needed to improve cognitive function. They also differ in that there are no fixed weights. Instead, they use dynamic weights that are determined by information-theoretic methods. Training with DANs is done by way of case-based reasoning. Additionally, new information can be added to DANs without re-training the network, a genuine advantage over more traditional models.

The CSML has a long history of different kind of projects. It began in the mid-1990's as the "Internet Applications Laboratory" developing search engines for academic use. As the "Digital Humanities" grew as an area of study, the lab's name was changed to the "Digital Humanities Lab," and then, recently, the "Cognitive Science Modeling Lab," to reflect more closely what we do here. Over the years, the lab has been staffed by more than fifty students working on a variety of projects, including internet search engine design, agent-based exploration of traffic light patterns in Evansville, Indiana and classroom simulations, and, regarding DANs in particular, on a range of models that include 1) object identification based on properties and context-sensitivity, 2) comparison of similarities and differences among properties and objects, 3) shape recognition of simple shapes regardless of where they might appear in an artificial visual field, 4) association across simulated sense modalities, 5) primary sequential memory of any seven digit number, 6) network branching from one subnet to another based on the presence of a single stimulus, 7) eight-bit register control that could perform standard, machine-level operations as with standard Turing-style computational devices, and 8) rudimentary natural language processing based on a stimulus/response (i.e. anti-Chomskian) conception of language.

After a decade of exploration in toy environments, we at the CSML stepped out into a genuinely complex adaptive system in a real-world stochastic environment, namely, the National Football League, where we are attempting to predict winners and losers of games and, in time, also the point spread. The NFL was chosen because, while it is massively complex, it is relatively constrained by a regular schedule, a fixed number of teams, a set of articulated rules, and regular stop points (unlike free-flow games such as hockey, basketball and soccer) where data can be discretely retrieved. 

During the 2014-2015 season, we employed our first model based on weighted averages determined by the percentage of points earned by a team during 2014-2015 season play only. As expected, the network learned as it went, starting in the 30 percentile for week two, then 50 percentile for week three, then on to end the season averaging around 65% correct when predicting winners. During two weeks, the model hit into the 80 percentile, but it also fell to the 30 percentile for two weeks as well.

This season, 2015-2016, we are moving into genuine network models based on our regular DAN methods described above. Pattern matching data will be based on drive profiles ranging over all the NFL games played over the last ten years. The method, however, will not permit robust predictions of the current season until each team has played at least one home and away game. In the meantime, we will roll out predictions based on last year's method. Our predictions and results will be regularly posted on this blog throughout the season.
 properties and context-sensitivity, 2) comparison of simi-larities and differences among properties and objects, 3)shape recognition of simple shapes regardless of wherethey might appear in an artificial visual field, 4) associationacross simulated sense modalities, 5) primary sequentialmemory of any seven digit number (inspired by Allen andLange 1995), 6) network branching from one subnet toanother based on the presence of a single stimulus, 7)eight-bit register control that could perform standard, ma-chine-level operations as with standard Turing-style com- putational devices, and 8) rudimentary natural language processing based on stimulus/response (i.e. anti-Chomskian) conception of language.

No comments:

Post a Comment