Table 5.

The difference between the $AI{C}_i^m$ score for the standard Q-learning network model described above and the various extensions of this model to include eligibility traces. The first row shows the improvement in fit quality when state cues are combined with the decaying action traces proposed by Bocagz, et al. (2007). The second row shows the improvement in fit quality using decaying state-action traces associated with previous inputs to the network (consistent with the Q(λ) algorithm of Watkins, 1989). As in the previous tables, positive values indicate conditions where the extended model provided a better fit on average than did the more restricted model controlling for the additional parameters. In parentheses is the percentage of participants for whom the extended model provides a better fit than the original.

	Experiment 1
Types of Traces	shuffled-cue	consistent-cue
Action Only	11.4 (0.94)	8.8 (0.94)
State-Action	6.8 (0.65)	4.1 (0.59)