The aim of this project is to build on existing relationships between certain types of market making and machine learning in order to develop new ways to compose predictive technology.
The kind of questions I hope to answer include:
To what extend can pricing or other market mechanisms be interpreted as learning algorithms.
Can well-understood learning algorithms be used to create novel market mechanisms?
How might we use market mechanisms to combine a variety of human and machine predictions to solve complex problem involving heterogeneous data?
Much of the work will be on the theoretical foundations of markets and learning. This is driven by a number of striking similarities between the mathematical formalisms used to describe and analysis both cost-function-based market making for prediction markets and certain online learning algorithms.
Reid, M.D., Frongillo, R.M., Williamson, R.C., Mehta, N., _Generalized Mixability via Entropic Duality, COLT 2015.
Frongillo, R.M. and Reid, M.D., Randomized Subspace Descent, Workshop on Optimization for Machine Learning, NIPS 2014.
Frongillo, R.M. and Reid, M.D., Risk Dynamics in Trade Networks, Workshop on Transactional Machine Learning and E-Commerce, NIPS 2014.
Premachandra, F.H.A.M., Prediction Markets for Machine Learning: Equilibrium Behaviour through Sequential Markets, Ph.D. thesis, The Australian National University, September 2014.
van Rooyen, B. and Reid, M.D., Conjugate Priors for Generalized MaxEnt Families, Proc. of MaxEnt, 2013.
Frongillo, R.M. and Reid, M.D., Convex Foundations for Generalized MaxEnt Models, Proc. of MaxEnt, 2013.
Premachandra, M. and Reid, M.D. Aggregating Predictions via Sequential Mini-Trading, Proc. of ACML, 2013.
Some related work that was developed before the project officially started includes:
This project is funding by a Discovery Early Career Researcher Award (DECRA) (DE130101605) awarded by the Australian Research Council to run from April 2013 through to March 2017.