Conférence d’Alexandre Hains, étudiant à la maîtrise en génie électrique

A problem resides in the fact that online modeling an unknown agent can be time exhaustive. Especially when data is available only in limited quantity or if the classes are heavily unbalanced during data acquisition; some classes are underrepresented because of a lower likeliness to encounter specific cases or situations.

In either scenario, the time required to acquire enough data from each class, in order to properly model the agent behavior, is increased. Without any prior, this modeling will take time. However, in a situation where a few already trained experts models similar to the agent are available, how can the learning time of a new agent be reduced? Time is of the essence  – we want to model a new expert’s behavior in a few steps (questions/answers) as possible.