Research on online experiment design (ABS):
One of the key chanlleges for machine learning is to obtain high quality labels for data, which can be expensive and/or noisy. We consider two popular methods for online experiment design with limit labels: Active learning uses computational algorithms to suggest the most useful measurements, and one possible theoretical framework is by balancing exploration and exploitation in a function maximization setting.
acton experiments: experiments for testing on acton package, which is an active learning Python library.
Multi-Armed Bandit (MAB) problems are well-studied sequential experiment design problems where an agent adaptively choose one option among several actions with the goal of maximizing the profit (i.e. minimizing the regret).
Upper confidence bounds: Among all polices have been proposed for stochastic bandits with finitely many arms, a particular family called "upper confidence bound" algorithm has raised a strong interest. The upper confidence bound algorithm is based on the principle of optimism in face of uncertainty.
Thompson sampling experiment: Gaussian reward with 5-arm bandit simulator.
One application for active learning and bandits algorithms is the experiment design for synthetic biology. The goal is to use machine learning algorithms (e.g. GPUCB) to identify the most probable combination of ribosome-binding site (RBS) which gives the best protein we need.
matrix factorization using MovieLens; tutorial for maxtrix factorization.