The Rescorla-Wagner Model

Suppose an animal has already learned that one cue predicts an outcome. Later, that familiar cue appears together with a new cue, and the same outcome occurs. Will the new cue automatically become a strong predictor? The Rescorla-Wagner model says that the answer depends on what remains surprising. If the familiar cue already accounts for the outcome, the compound produces little new prediction error. The new cue therefore receives little or no additional associative strength.

The Rescorla-Wagner model treats conditioning as error-correction learning. On each trial, an animal has a prediction based on previously learned associative strengths. The actual outcome is then compared with that prediction. The difference between them is prediction error. That mismatch determines how much further updating is justified. A large mismatch provides a basis for substantial change. When the prediction already accounts for the actual outcome, the error can be zero or low, so little additional learning occurs.

Consider a simple generated scenario. At the beginning, a cue has little or no prior conditioning history. An outcome occurs after the cue, so the outcome is not yet fully accounted for by the cue's prediction. The mismatch creates a basis for increasing the cue's associative strength. On later trials, the cue contributes to a stronger prediction. If the same outcome continues to occur, the remaining error becomes smaller, and the amount of further learning becomes smaller as well.

Kamin blocking is a two-stage conditioning situation. In the first stage, one stimulus is repeatedly involved in conditioning with an outcome. In the second stage, that previously conditioned stimulus appears together with a new stimulus, followed by the same outcome. The new stimulus is the added component of the compound. According to the Rescorla-Wagner explanation, the prior cue already predicts the outcome. The compound therefore introduces no new surprise, or only a small amount. The prediction error is zero or low, so the new component gains little or no additional associative strength.

The connection to the Least Mean Square learning rule is conceptual and structural. In the Rescorla-Wagner model, associative strengths are adjusted to reduce prediction error. In LMS learning, weights are adjusted to reduce error. Both accounts therefore make the error signal central: the current prediction is compared with what actually occurs, and the adjustable quantities are changed in response. The source does not provide a formula or numerical update rule here, so the comparison should be understood as a correspondence between learning ideas, not as an equation to calculate.

A central idea associated with the Rescorla-Wagner model is that animals learn when they are surprised. This emphasis on surprise helps explain why adding a stimulus does not always produce new learning. If earlier conditioning has already prepared the animal for what will happen, the added stimulus may contribute little or no additional learning. Kamin blocking provided an important historical basis for this way of thinking.

The word reinforcement has changed meaning across historical traditions in animal learning. According to the source, the term first appeared, to the best of its knowledge, in the 1927 English translation of Pavlov's monograph. In that use, it referred to a non-action-contingent case, meaning that the event was not described as depending on an action. Later, Mackintosh proposed using reinforcement for either strengthening or weakening a pattern of behavior. Skinner used the term only for strengthening behavior and treated weakening as something produced by punishment. These differences matter because the same word can point to different ideas in different traditions.