FCAI MLCS - Computational Rationality

Title : FCAI MLCS - Computational Rationality Author(s): Andrew Howes

• Some methods in which humans act can be modelled as actions taken under bounded computational resources, e.g. breaking phone numbers into chunks tto remember it, or adaptation to motor uncertainty e.g. where to tap to open an app in your phone, where users will aim to overshoot to the edge to avoid pressing the neighbouring app, or when seeking new information when comparing things e.g. eye-tracking data in apartment listing experiment shows users comparing same features between samples if information is presented at once, but if the user has to click to view a feature they press all features of a given sample first, and their choice of apartment may be different, or in adaptive planning e.g. Tetris if the cost of moving and rotating the blocks is cheap people would rather try it in the game (epistemic action) rather than thinking about it in their mind or in a 2048 style game if there is no friction between the tiles users simply try to move the blocks around to see what happens but if their is a delay i.e. friction when moving the tiles they reduce the epistemic action and do more internal planning, and this thinking enables them to view the game representation as a sequence of tiles rather than individual tiles.
• Computational rationality states that cognition can be specified as an optimal program problem defined by an adaptation environment, a bounded machine and a utility function. This formalization is adapted from Provably bounded optimal agents by Russel, Subramanian. Hence the ability to reason is bounded by a physical machine and utility function, both of which are evolving, within the constraints of the environment.
• Probability theory plus noise explains biases injudgment presents an explanation of experiments by Kahneman and Tversky, where they claim probability theory and a machine which samples with noise from memory can explain many cognitive biases like the conjunctive fallacy, over/under confidence, subadditivity, i.e. people are surprisingly good at following probability theory given the bounds imposed by the machine and environment.
• Biases, seen as suboptimality should be a sign of an absence of an explanation rather than an explanation.
• Intrinsic motivation is an unexplored topic from a computational perspective, we do not know anything about these utility functions.
• The theory of computational rationality is in contrast to the view that people are irrational, biased and need nudging to behave rationally.

Can any of these be related to Interactive Proof systems? Can reasoning with probabilities in this case be done using self-referential probability?

Also I think there was a paper by Dayan on modelling computationally rational humans as POMDPs or maybe I misheard it.