Amir Sani, Philosopher dreaming on the edge of the convex hull
Thank you for checking out my page. My name is Amir Sani and I'm a member of the SequeL Team at INRIA-Lille Nord-Europe. I'm a Machine Learning doctoral student supervised by Rémi Munos and Alessandro Lazaric researching online learning algorithms and the dimension of risk aversion on decision making in machine learning. During my PhD, I've contributed the following work to this end:
  • Exploiting Easy Data in Online Optimization:
    • This is a very interesting algorithm that provides a meta-framework by which one can lower bound the performance of full information online learning with a benchmark, which can take the form of an existing algorithm, learning rate which favours a better lower bound or non-selectable expert. 
    • This algorithm works well in scheduling simulated workloads on heterogeneous compute environments, online portfolio selection, portfolio insurance, portfolio replication and as a regularizer.
  • Risk aversion in multi-arm bandits:
    • This work is interesting from a risk perspective because it captures both variance aversion and regret aversion. More specifically, the UCB bound and realized regret can be seen as adjustments akin to the regret adjustment within the Regret Theoretic Expected Utility theory of Bell, Loomes and Sudgen.
    • This algorithm works well in settings such as a decision set defined by the optimal limit order distance (Avellaneda and Stoikov).
I've also spent a considerable amount of time and thought on realistic simulation of dependent time processes through bootstrapping dependent sequences. This work is model free and fairly automatic. This work along with software wrappers in Matlab, R and Python are forthcoming.

More calculation is better than less, Some calculation is better than none
-Tzu, Sun (6th century BC), The Art of War