Pedro Salcedo Lagos psalcedo@udec.cl MISTRAL: A Knowledge-Based System for Distance Education that Incorporates Neural Networks Techniques for Teaching Decisions Pedro Salcedo Lagos psalcedo@udec.cl Universidad de Concepción, Chile
Nobel Prize laureate in Literature (Chile), 1945 Gabriela Mistral Nobel Prize laureate in Literature (Chile), 1945
This is the place where I come from
Knowledge Based System Motivation Neural Networks MISTRAL Knowledge Based System Internet Distance Education
Contents Current Platforms (Overview) The Platform MISTRAL NNs in MISTRAL
Some revised Platforms The Learning Manager WebMentor Alfa Net - UNED IVLE LUVIT Asymetrix Librarian Virtual-U eduprise.com Blackboard Convene Embanet eCollege.com IntraLearn Symposium TopClass WebCT
Tools for Instructors Course planning On-line evaluation Tools for Instructors Course planning Most of platforms On-line evaluation Some of platforms Specific material addressing A few platforms
Instruction Advantages Consistent Interface Most of platforms Synchronic communication Support internal e-mail Some platforms Help on-line A few platforms
Tools for Students Multimedia support Discussion groups Most of platforms Discussion groups Anouncement panel Some platforms Progress tracking A few platforms
And Costs... Some are free (AulaWeb) Some are expensive (about US $ 90,000 for Docent) Some have a cost per year (about US $ 5,000 for Course Info) Some include maintenance costs..., creation costs...
Most of them considering 5 tasks Material preparation Teaching-learning strategies generation Dialog capability Evaluation Management
Material Preparation Current Proposed by MISTRAL Time consuming Use activities instead of contents Examples: Read chapter X in text Y Use simulation tools Participate in a discussion group
Strategies Generation Current Found in only one commercial platform (Cameleon) Proposed by MISTRAL Adapts to student profile considering learning styles and previous knowledge Risk students’ detection
Dialog Capability Current Proposed by MISTRAL Exists in most of platform Without using AI techniques Proposed by MISTRAL Adapts media to user profiles using AI rules Automatic detection of interaction degree Personalized activities (via e-mail, portfolio)
Evaluation Current Proposed by MISTRAL Low level of development Assumed by the instructor Proposed by MISTRAL Adaptive evaluation based on user modelling use of Bayes nets Portfolio techniques use of virtual folders
Management Current Proposed by MISTRAL Exists in all revised platforms Lack of support for instructor or manager Proposed by MISTRAL Automatic statistics generation general course behavior interactivity level ... via e-mail
MISTRAL Adapts to students’ profile by suggesting Activities to be accomplished Tools to be used Evaluation mechanisms for correcting detected problems
MISTRAL (cont) Multiuser modality Easy knowledge acquisition contents and activities for each course Folder structure based in portfolio strategy Diagnostic and detection of knowledge level using rules and Bayes nets
MISTRAL (cont) Virtual room personalized profile group interaction warning messages when participation level of students is lower than expected
Neural Networks in MISTRAL Rules Objects Specified knowledge Corrective actions generation when detecting possible unsuccessful performance
Sequence of activities in MISTRAL for NN use yes Get student Trained net Information no Detect Failure Net training Risk no Failure Probabil. yes Determine corrective actions
Application development The neural net has one hidden layer and number of nodes is variable depending on number of factors to be considered The activation function is hyperbolic tangent
If Neural Network option is used
There are three additional entries
Following with variables configuration
Training the net
Visualizing learning parameters
Net training During training, the net learns to classify students in successful group or failure risk group depending on different psicosocial characteristics upoin requirements Data can be classified as default data (provided by MISTRAL) or user provided data
Implementing
Technical considerations Windows 2000 server (translating to Linux) Database access through ODBC using SQL (translating to MySQL) Programming in VBScript, JavaScript and Java with ASP (translating to PHP) Bayes net has been programmed using JavaBayes library
Conclusions MISTRAL Platform allows increase productivity both in organizations oriented to distance education and in instructor work improvement in distance courses management maintain intellectual property get a better distance education teaching/learning process, really personalized and efficient
Current work Search for more adequate threshold functions, learning rules and NN architecture to determine students’ success Learning coefficient and momentum revision, more adequate to Generalized Delta Rule Net training to adapt learning strategies to different contexts Different courses knowledge acquisition
Thanks for your time... Pedro Salcedo Lagos