Rats can learn that cocaine and food reinforcers are smaller than expected: A test of a computational model of addiction

According to Redish's (2004) computational model of addiction, drugs of abuse differ from natural rewards (e.g., food) in that they produce a continual, non-correctable overexpectation for the reward. The present experiment tested the model-based prediction that subjects should be incapable of learning that a cocaine reward is smaller than expected, but should be able to learn that a food reward is smaller than expected. This was done by training rats to expect a large cocaine or food reward, subsequently presenting them with a smaller-than-expected reward, and then seeing if they could learn from this discrepancy. Results indicated that subjects were able to learn that a cocaine or food reward was smaller than expected. This outcome suggests that traditional error-correction learning models, where reward expectation decreases after receiving smaller-than-expected rewards instead of boundlessly increasing as it does in Redish's theory, apply to both drug and natural rewards.