Dandurand et al. have devised a computational model, and a way to train it, that better matches human performance on learning a certain problem-solving task than their previous model. The problem to be solved is to find which of 12 items is heavier or lighter than the others using a balance no more than three times. A standard temporal-difference machine learning algorithm is used. What is different is that extra terms have been added to reward the model for moving to a state that is closer to a solution state and to penalize the model for choosing item distributions on the balance that are complex or asymmetrical with respect to the information inherent in those distributions. The reported simulations and experiments with human subjects show that these added terms make the computational model behave more like humans. The discussion of why these extra terms are beneficial and why they still do not capture much of the cognitive processes involved in learning to solve problems is especially interesting to read.

The computational model contains a neural network used to store the action policy information; it takes the place of the state-action look-up table that would ordinarily be used in the machine learning algorithm. This use of neural networks, and this paper generally, will be of interest to anyone studying neural networks, cognitive modeling, problem solving, and machine learning. The paper is well written with lots of references to related work.