[1]See Lovett & Anderson (1995) for a discussion of how history of success and apparent appropriateness of various choices jointly influence selections in problem solving.

[2]Our modeling goals differ from those of related machine-learning work (e.g., Sutton, 1988).

[3]In ACT-R, the estimate of PPS can be influenced by other factors as well, but here it is relatively well modeled as a monotonic function of success rate alone.

[4]Note that the RHS of this equation does not include unsolvable problems but the LHS does. Thus, the equation reduces to probability matching when all problems are solvable (i.e., #A_Solutions + #B_Solutions = all trials).