Our research strives to connect three areas: Clustering, tomography and data analysis (or machine learning). Building onto established concepts from machine learning like kernel functions, we are currently developing a new idea for a multi-class learning algorithm whose key element is a new type of data transformation.
In this talk, we recap the inspirations for our new approach from the above three areas and formally introduce our data transformation and algorithm. Further, we present experimental results for some toy examples with an emphasis on the comparison to other, established learning techniques from machine learning.
Richard Littmann: "Competitive Equilibria in Combinatorial Exchanges with Financially Constrained Buyers"
Growth in air traffic has made airport capacity a very scarce resource and increases pressure to implement markets for the allocation of airport time slots. Package bids are essential in this domain and thus, combinatorial auctions and exchanges have been proposed as market mechanisms for time slot allocation. Although some slots might have a high Net Present Value to airlines, especially smaller airlines are possibly limited by their financial capacities. These budget constraints have mostly been neglected in combinatorial auctions, leading to inefficiencies such as bid shading. Since payments, and therefore also bids, may never exceed the budget, buyers are potentially hindered to communicate differences in values, too. To prevent these kinds of inefficiencies, we model budget constraints explicitly.
We introduce mixed integer bilevel linear programs (MIBLP) to compute the welfare maximizing allocation in the core, i.e. no subset of participants can achieve a better outcome by only trading among themselves. Despite the hardness of the problem, we present results on instances that are practically relevant and comment on several techniques we applied to make the problem more tractable.
Venue: TUM City Campus Room Z 532
Date: Monday, July 9th, 2018, 14:00