Periodic Reporting for period 3 - MIAMI (Machine Learning-based Market Design)
Período documentado: 2021-12-01 hasta 2023-05-31
Market designers study how to set the "rules of a marketplace" such that the market works well. This is important for society, such that resources are allocated efficiently (put to good use), such that market participants do not have an incentive to manipulate the market mechanism, and such that the marketplace produces a fair outcome.
However, markets are getting increasingly complex such that designing good market mechanisms "by hand" is often infeasible, in particular when certain design desiderata are in conflict with each other. Moreover, humans are boundedly-rational: already in small domains, they are best modeled as having incomplete preferences, because they may only know a ranking or the values of their top choices. In combinatorial domains (where participants make choices regarding bundles of options), the number of choices grows exponentially, such that it quickly becomes impossible for humans to report their full preferences to the market mechanism.
In this ERC project, we combine techniques from "machine learning" with "market design" to address these challenges. The main idea is that machine learning can help deal with incomplete data. For example, a machine learning algorithm can learn the preferences of a human in a marketplace from just a few observations, and these learned preference functions can then be used by a market mechanism to make better decisions. But this also introduces new challenges, as the learned preferences may be wrong, or the market participants may manipulate the machine learning algorithm on purpose. Thus, when designing machine learning-based market mechanism, we take these challenges into account.
In addition to pushing the scientific boundaries of market design research, this ERC project also has an immediate impact on practical market design. We apply our techniques in two different settings: (1) for the design of combinatorial spectrum auctions, a multi-billion dollar domain; and (2) for the design of matching markets (e.g. school choice, refugee matching, adoption matching, course allocation).
However, markets are getting increasingly complex such that designing good market mechanisms "by hand" is often infeasible, in particular when certain design desiderata are in conflict with each other. Moreover, humans are boundedly-rational: already in small domains, they are best modeled as having incomplete preferences, because they may only know a ranking or the values of their top choices. In combinatorial domains (where participants make choices regarding bundles of options), the number of choices grows exponentially, such that it quickly becomes impossible for humans to report their full preferences to the market mechanism.
In this ERC project, we combine techniques from "machine learning" with "market design" to address these challenges. The main idea is that machine learning can help deal with incomplete data. For example, a machine learning algorithm can learn the preferences of a human in a marketplace from just a few observations, and these learned preference functions can then be used by a market mechanism to make better decisions. But this also introduces new challenges, as the learned preferences may be wrong, or the market participants may manipulate the machine learning algorithm on purpose. Thus, when designing machine learning-based market mechanism, we take these challenges into account.
In addition to pushing the scientific boundaries of market design research, this ERC project also has an immediate impact on practical market design. We apply our techniques in two different settings: (1) for the design of combinatorial spectrum auctions, a multi-billion dollar domain; and (2) for the design of matching markets (e.g. school choice, refugee matching, adoption matching, course allocation).
Regarding markets with money: (a) we have developed a new machine learning-powered iterative combinatorial auction, using support vector regression, which outperforms the auction design currently used in practice in terms of efficiency; (b) we have developed the first machine learning-powered combinatorial auction with interval bidding (allowing bidders to submit upper and lower bounds instead of exact values), which reduces the cognitive efforts for the bidders in an auction (Beyeler, Brero, Lubin, Seuken, EC'21); (c) we have developed the first deep learning-powered combinatorial auction (Weissteiner and Seuken, AAAI’2020); (d) we have developed a new machine learning framework to capture the uncertainty of the predictions of a neural network; (e) we have designed a novel machine learning-based combinatorial auction that combines techniques from Fourier analysis with deep neural networks; (f) we have designed a novel neural network that is particularly well suited to learn agents' preferences in combinatorial assignment domains; (g) we have started a new research project on market design for drone traffic allocation (Seuken, Friedrich, Dierks, AAAI'22).
Regarding markets without money: (a) we have developed a machine learning-based matching algorithm for refugee allocation that enables making trade-offs between the interests of the refugees and the interests of the host countries; (b) we have started a research project on a machine learning-based adoption matching mechanism; (c) we have started a new research project on machine learning-based course allocation.
Regarding markets without money: (a) we have developed a machine learning-based matching algorithm for refugee allocation that enables making trade-offs between the interests of the refugees and the interests of the host countries; (b) we have started a research project on a machine learning-based adoption matching mechanism; (c) we have started a new research project on machine learning-based course allocation.
Our work on machine learning-powered combinatorial auctions has shown, for the first time, that incorporating machine learning algorithms into an iterative combinatorial auction can lead to significant efficiency improvements. Already now, in large domains, our designs outperform the combinatorial clock auction (CCA), the auction design most often used in practice. Until the end of the project, we expect that we will improve our techniques further, achieving even higher efficiency with lower cognitive costs for the bidders. One promising technique in this regard is the uncertainty quantification for neural networks we are currently working on. With this technique in hand, we will be able to develop Bayesian Optimization techniques that can be incorporated into an iterative combinatorial auction, which we expect will increase the performance of our designs even further and which should also pave the way for many other exciting applications.
In our work on matching markets, we have developed the first model to study adoption matching mechanisms from a market design perspective. Until the end of the project, we expect that we will be able to compare two of the main approaches how adoption matchings are currently performed regarding their social welfare and also develop the first machine learning-supported adoption mechanism.
In our work on course allocation, we have developed a new machine learning-based course allocation mechanism, which, in some setting, outperforms the current state-of-the-art mechanism (CourseMatch) in terms of allocative efficiency. We are currently work on further extending this work, with more sophisticated preference elicitation paradigms. Until the end of this project, I expect that we will be able to significantly reduce the elicitation costs for the agents while increasing overall efficiency.
In our work on matching markets, we have developed the first model to study adoption matching mechanisms from a market design perspective. Until the end of the project, we expect that we will be able to compare two of the main approaches how adoption matchings are currently performed regarding their social welfare and also develop the first machine learning-supported adoption mechanism.
In our work on course allocation, we have developed a new machine learning-based course allocation mechanism, which, in some setting, outperforms the current state-of-the-art mechanism (CourseMatch) in terms of allocative efficiency. We are currently work on further extending this work, with more sophisticated preference elicitation paradigms. Until the end of this project, I expect that we will be able to significantly reduce the elicitation costs for the agents while increasing overall efficiency.