Ranking simulated policies under deep uncertainty
Either due to lack of data, knowledge or high sensitivity to some free parameters, simulations often produce large prediction intervals where “almost anything can happen”. The standard approach to tame this problem is to look at some “average” run or behaviour and perhaps to rank policies by their average performance. Even when prediction intervals are large, these averages tend to converge quickly giving the impression that such rankings are both easy to produce and informative. I would like to discuss how these averages are actually uninformative when, as it surely almost the case, the priors on parameters are not well thought out (as for example using a uniform prior as a substitute for lack of knowledge). Rather, the best approach should be to perform a thorough parameter space exploration and cluster the problem into separate scenarios, identify the key parameters dividing them and exploiting simulation’s cues to produce better policies.
Date: 14 November 2019, 15:00 (Thursday, 5th week, Michaelmas 2019)
Venue: Manor Road Building, Manor Road OX1 3UQ
Venue Details: Seminar room A
Speaker: Ernesto Carella (Oxford Martin School, Oxford University)
Organising department: Institute for New Economic Thinking
Organiser: Susan Mousley (INET Oxford Admin Team)
Organiser contact email address: info@inet.ox.ac.uk
Booking required?: Not required
Audience: Members of the University only
Editor: Susan Mousley