Classical model of action:

  • Optimal action depends on state of the world
  • Therefore, first step of action is to (1) form a belief (analyse surroundings/prospects)
  • (2) imagine a value function of next state brought about by action
  • (3) optimise action that maximises value of the next state

Model of action

  • Classical model doesn’t work when the best next thing to do is to search for/resolve uncertainty
  • Optimal action depends on beliefs about the world, and subsequent action
  • Further, it’s a function of the order in which you interrogate the world
  • Therefore the functional (function of a function) to be optimised is a function of beliefs
  • Optimal action therefore is optimising sequences or policies of actions
  • To be optimised: a function of a belief, integrated over time

Free Energy Principle:

  • The goal of a self-organising (eg biological) system is to minimise prediction error (surprise), also called ‘free energy’, by forming continually-updated beliefs/inferences about the world from which to form policies of action
  • Friston considers this an organising principle of all life and intelligence
  • To be alive (to be a system that resists disorder and dissolution) is to act in ways that reduce the gulf between your expectations and your sensory inputs (AKA, to minimise free energy)

  • If a prototypical agent, or a ‘good agent’ minimises free energy (thereby minimising ‘surprise’), they must believe that the actions they take minimised expected free energy
  • expected free energy associated with a policy of action is minimised

Markov Blanket:

The Markov Blanket is a concept in machine learning which is essentially a shield that separates one set of variables from others in a layered, hierarchical system. The blanket defines the boundaries of a given system. That is, in cognition, a cognitive version of a cell membrane shielding states inside the blanket from states outside. This is the schema by which surprise is minimised— the Markov blanket is a set of variables sufficiently complete that another random variable can be inferred from it . If a Markov blanket is minimal (parsimonious) (cannot drop any variable without losing information), it is called a Markov boundary.