Thinking Fast and Slow in AI: the Role of Metacognition

gist

Bergamaschi Ganapini et al. propose SOFAI: System 1 solvers answer from experience by default, and a metacognitive agent decides when to call System 2. Fast solvers run in constant time from a model of self; slow solvers reason only when invoked. A cheap MC1 stage accepts a confident System 1 answer or insufficient resources; MC2 activates System 2 only when expected extra reward exceeds expected cost. Ongoing instances target constrained grid path finding and epistemic planning.

ideas

• SOFAI is System 1 by default. Incoming tasks go first to constant-time experience-based solvers; System 2 agents never run unless the metacognitive module invokes them.
• Models of self, world, and others support both solver kinds. A background updater keeps past decisions, resource costs, rewards, environment knowledge, and beliefs about other agents current.
• MC1 often stops at System 1. It accepts the fast answer when confidence is high relative to expected reward, or when there are not enough resources for a deeper check.
• MC2 pays for System 2 only when the extra expected reward beats expected cost. It compares the System 1 action's expected value against the cost of running a slower solver.
• Sequential problems can arbitrate per step or once per plan. Per-decision metacognition is more flexible; whole-plan solvers can use deeper domain knowledge when metacognition runs once.

quotes