See sec 14.2.2 in Sutton-Barto

Rescorla and Wagner created their model mainly to account for blocking. The core idea of the Rescorla–Wagner model is that an animal only learns when events violate its expectations, in other words, only when the animal is surprised (although with- out necessarily implying any conscious expectation or emotion).

It can be interpreted as Stochastic gradient descent to minimize the expected square error of a linear model between stimuli and responses (where the true response is determined by the unconditioned stimulus)