Computer Can Now Predict Learning Patterns

New computer model shows great anticipatory skills



Computers have come a long way from closet-sized calculation machines



Scientists from Penn State's College of Information Sciences and Technology managed to create the first computer
model predictive of human action. The program, named Soar, was able to
foretell accurately the times 10 students needed to complete basic
trouble-shooting tasks, such as identifying misfit components inside a
circuitry.




Naturally, after solving a few tests, the task became easier for
humans, as they learned and evolved their abilities to identify and
solve problems inside the circuits. The program also evolved and
figured out that people would take less time to spot similar mistakes
in similar circuits. Thus, Soar was able to predict the participants' times with a margin of 2 to 4 seconds.




In addition, in 8 out of 10 cases, it came extremely close to
identifying the exact times the test subjects needed to complete the
tasks. "The model does not merely accurately predict
problem-solving time for the human participants; it also replicates the
strategy that human participants use, and it learns at the same rate at
which the participants learn,” explained IST and psychology associate
professor, Frank Ritter.




"The project shows we can predict human learning
on a fine-grained level. Everyone thinks that's possible, but here's an
actual model doing it. The model provides a detailed representation of
how a transfer works, and that transfer process is really what
education is about," he added.




The fact that the Soar program was able to adapt to changes and learn
progressively brings new hopes for informatics experts to create more
and more advanced software, which could eventually replace humans in
dangerous activities. Also, progressive learning is the key for smarter Artificial Intelligences that could learn to adapt to challenges and solve problems more efficiently.



Penn scientists worked with University of Nottingham researcher Peter
Bibby and two of his research assistants to optimize Soar. The
Cognitive Science journal published their study, “Modeling How, When,
and What Is Learned in a Simple Fault-Finding Task,” in the July/August
2008 issue.