Software Project Management Plan

2.2.1 Teams and Tasks - Learning Subsystem

• Be able to deal with a large number of users (over 6000 total dealers)
• Investigate mechanisms such as machine learning algorithms and agent technologies
• Investigate existing monitoring and diagnostics tools
• Provision of reports detailing the network activities with a graphical user interface
• Streamline the information transaction process based on user activity patterns

2.32 Meeting Times - Learning Subsystem

2.4 Project Responsibilities - Learning Subsystem

3.2.1 Assumptions - Learning Subsystem

• Learning is on server side
  
• Different types of data to analyze; one general system
• Using 3rd party software whenever possible
• Rely on database triggers for actions

3.2.2 Dependencies - Learning Subsystem

• Rely on database team for persistent storage (log, behavior files)
• Rely on database for actions

3.2.3 Constraints - Learning Subsystem

• Learning problem essentially data mining
• Everything needs to be Java or wrapped in Java
• Domain specific knowledge may be limited

3.3 Risk Management - Learning Subsystem

• Risk: Code could be compute intensive
• Contingency: Fundamentally, the learning system should be written with fast algoritms and fast 3rd party code. The data mining process will be performed off-line, since it will be the most costly of all the learning operations. Finally, to avoid increased iterations, there will be a default, less intelligent behavior, that the system will recommend when appropriate.
  
  
• Risk: No ability to use 3rd party code
• Contingency: A decision tree algorithm will yield a suboptimal solution for all the scenarios and will take a relatively shorter time to implement. The learning system, therefore, will not overapply the various learning techniques available. When looking for 3rd party code, however, we will not search for too much domain specific code; but rather we will try to modify the code we find.
  
  
• Risk: No 3rd party packages in Java available
• Contingency: In this case we will write Java "wrappers" to implement code written in other languages. We can also obtain source code for packages that enable cross-compiling.
  
  
• Risk: Need more than 1 learning algorithm for optimal performance
• Contingency: Here there are two available choices. We could either use multiple learning methods (most likely using multiple 3rd party packages). Otherwise, we must accept a suboptimal but nonetheless effective single solution.
  
  
• Risk: Assumptions change because of other teams assumptions
• Contingency: This problem can be resolved in the systems design phase.
  
  
• Risk: Subsystem assumptions invalidated
• Contingency: We will start from the beginning and revise our plans.