Abstract: 

In recent years, inducing a desired behaviour in learning agents by exercising some degree of influence over their experiences has received extensive treatment. While there are several offshoots of this problem, it is commonly viewed within the framework of Curriculum Learning: devising a sequence of tasks that gradually induce the desired behaviour. Majority of these methods assume value alignment between the teacher and the learning agent, though some notable exceptions exist. One such exception is the method of Behaviour Cultivation (BC) that induces a desired behaviour by tweaking the environment dynamics.

In this talk, the original BC will be introduced, noting that it does not assume value alignment, and that it is indispensable, i.e. not reproducible by other teaching methods. This will be followed by recent steps taken to address BC's weaknesses -- being an open loop methods, and lack of consideration to a deployment in a group of agents. The talk will conclude with the outlook at future research steps and the approaching applications horizon.

Bio:

Zinovi Rabinovich is currently an Assistant Professor at the Nanyang Technological University (NTU), which he has joined in 2017 after holding an industrial position for 5 years as a Senior Algorithms Engineer at Mobileye Vision Technologies Ltd. His sustainedpublication record highlights two key directions of his current research agenda: social behaviour analysis; and choice manipulation. The latter, applied to Security Games via strategic information disclosure, was published in several seed papers in 2015. It has sincegathered an independent following. The academic record of Dr. Rabinovich also includes post-doctoral appointments at the University of Southampton (2007-2010) and the BarIlan University (2010-2011). His PhD is from the Hebrew University of Jerusalem (2008).