Added: Jul 26, 2008

From: googletechtalks

Duration: 48:9

Google Tech TalksMarch, 14 2008ABSTRACTThe online learning framework captures a wide variety of learning problems. The setting is as follows - in each round, we have to choose a point from some fixed convex domain. Then, we are presented a convex loss function, according to which we incur a loss. The loss over T rounds is simply the sum of all the losses. The aim of most online learning algorithm is to minimize *regret* : the difference of the algorithm's loss and the loss of the best fixed decision in hindsight. Unfortunately, in situations where the loss function may vary a lot, the regret is not a good measure of performance. We define *adaptive regret*, a notion that is a much better measure of how well our algorithm is adapting to the changing loss functions. We provide a procedure that converts any standard low-regret algorithm to one that provides low adaptive regret. We use an interesting mix of techniques, and use streaming ideas to make our algorithm efficient. This technique can be applied in many scenarios, such as portfolio management, online shortest paths, and the tree update problem, to name a few.Speaker: Seshadhri Comandur - Research Scientist - New Grad - Mountain View

Channel: People

Tags: education  engedu  google  googletechtalks  talk  talks  techtalk  techtalks 

Rating: 4.25 (8 ratings)    Views: 3253' favoriteCount='21    Comments: 1

pav930t Says:

Jul 26, 2008 - Can you post some results where you have used the online learning framework for a portfolio management scenario? Any results that show performance compared to other approaches such as a Genetic Algorithm (where a gene represents a share in your portfolio chromosome), PGA, or others... Convergence results?