Rats repeatedly acquired the performance of selecting only the four baited arms in an automated eight-arm radial maze, with the arms containing food pellets randomly assigned prior to each session. During each 14-trial (trial: obtain all four pellets) daily session, the number of errors (selecting nonbaited arms or repeating arm selections) showed a within-session decline, and choice accuracy for the first four arm selections showed a positive acceleration across trials for all rats. An index-of-curvature statistic, calculated for total errors, was used to quantify both the within- and between-session improvement of performance. Scopolamine (0.03 to 0.3 mg/kg, ip), but not methylscopolamine (0.3 mg/kg), reduced the accuracy of the first four selections of each trial and increased total within-session errors for all rats. Session times also were increased by scopolamine. An examination of within-session accuracy showed only slight signs of improvement at the higher dosages of scopolamine. The results indicate that behavior in transition states maintained by reinforcement contingencies in the radial maze is similar to that maintained by extended chained schedules, despite the fact that some of the stimuli controlling behavior in the maze are absent at the moment behavior is emitted.