⁸It may seem logically contradictory to predict an RPE, or to predict a prediction error. If the agent were able to predict a prediction error, wouldn’t it mean that the agent understands what kind of error is about to occur, and then it should be able to cancel it by improving the prediction accordingly? This would indeed be true in the basic case where the prediction is only about a single quantity such as the expectation. If the agent somehow understands that it is predicting the expected reward as too high, it can simply make its prediction a bit lower, and thus the error in the prediction of the expected reward can be removed. However, this is no longer meaningful with more sophisticated predictions that predict the whole probability distribution. If the agent predicts a large risk, it predicts a major possibility of prediction error, but there is nothing wrong with that prediction; there is nothing to correct. Therefore, there is no contradiction in predicting that there is going to be a prediction error. Dabney et al. (2020) claim that the brain is coding the whole distribution of RPE, based on a populations of neurons with different thresholds.