At our next research meeting, Michael K. Cohen will give a talk on reinforcement learning with KL-regularization to a trusted policy:
Reinforcement learning, but don’t do anything I wouldn’t do
Cohen is a postdoc at UC Berkeley’s Center for Human Compatible AI. His papers and (highly recommended) blog “Solutions in Theory” argue for the theoretical safety of carefully designed agent architectures, often inspired by AIXI or leveraging intuitions from algorithmic information theory.
We will use the following NEW zoom link for the meeting, at the regular time (4 pm EST = 9 pm BST) on Monday, Sept. 22nd:
https://uwaterloo.zoom.us/j/7921763961?pwd=TDatET6CBu47o4TxyNn9ccL2Ia8HN4.1
However, some attendees (e.g. those in the UK) find the research meetings a little bit late in the evening, so I have created a when2meet to select a new time for future meetings (starting with October 6th, NOT this one), so please fill in your availability below if you plan on attending:
Universal Algorithmic Intelligence 
Leave a comment