8.6 Novelty, surprise, and interestingness

Barto, Mirolli, and Baldassarre distinguish novelty relative to memory from surprise relative to prediction [ Barto et al. , 2013 ] . Both can guide exploration, but neither is sufficient for scientific value. A novel state may be irrelevant; a surprising result may expose sensor error; and an important result may become unsurprising once expressed in the right coordinates.

Exploratory systems should therefore maintain several observers rather than a single creativity score:

Validity

asks whether the construction belongs to the declared space.

Novelty

compares it with the retained repertoire and equivalence rules.

Surprise

compares observation with prediction.

Structural yield

measures new consequences, factorizations, or distinctions made available by the result.

Transfer

asks whether the discovered path or object helps in a second problem.

Palm’s information-theoretic treatment is useful for the novelty and surprise observers [ Palm , 2012 ] . The other observers prevent an information measure from silently becoming a complete theory of creativity.