Unsupervised Heuristic Decision Making

Heuristic decision making, that is, decision making based on approximate reasoning, is a very important part of our cognition and humans are capable of figuring out effective heuristics for a rather wide palette of decision problems. Heuristics are also heavily used by machines to guide the search processes of solving NP-hard problems of interest. Unlike humans, however, the machines were not expected to "figure out" a heuristic for a given problem, but just to use the heuristics provided by human operators. Considering this situation, a natural (and economically very important) question is thus whether we can possibly build machines capable of automatically devising heuristics for problems they have never seen before? For years, the prospects of such fully unsupervised heuristic decision making were discussed in the literature with a heavy pessimism.

Over the last two decades, however, things changed dramatically, taking unsupervised heuristic decision making from being a crazy dream to advancing the state of the art in industrial autonomous systems. In fact, with the more complicated decision problems, the machines appear to be as effective as human "heuristic engineers", and sometimes much more effective! In this talk, we'll start with a (very speedy but reasonably comprehensive) overview of two seemingly unrelated concepts: heuristic decision making and model-based problem solving. We then connect between these two concepts, diving into the key representational and algorithmic ingredients (and success stories) of unsupervised heuristic decision making. In particular, along this journey I hope to show you that research in this area can be both mathematically/algorithmically rigorous, and empirically fun, and valuable in practice.