Fig. 4: Results of behavioral and decoding analyses based on the computational model of human navigation.

a Examples of subjects’ actual behaviors in the maze (left panels) and estimated behaviors using a hidden Markov model (HMM) with the maximum a posteriori probability estimate (right panels). Subject 1 and Subject 2 are representative examples of good and poor performances, respectively. The arrows represent the actual (black) and estimated paths (blue), starting from unknown initial positions (filled circles). The dashed lines on the estimated path indicate the subject’s position before and after the HMM inferred that the subject had re-estimated their position due to a discrepancy between the expected and observed scenes. The circle or square markers signify the positions in which the subject made a correct or incorrect scene prediction, respectively, and the color corresponds to the subject’s reported confidence level about the scene prediction in that position (green: low, orange: high). The color of the cross-markers in the right panels represents the confidence level about the current state estimated by the HMM (green: low, orange: high). b Proportion of trials with a high state-confidence level as a function of the prediction trial number in each game. The proportion of the trials in which the HMM estimated the subjects’ confidence level as high increased as the number of prediction trials performed in a single game increased (r = 0.63, p = 2.6 × 10⁻²⁸). Each box extends from the lower to upper quartiles with a horizontal line at the median. The whiskers represent 1.5 × IQR. The cross-markers indicate the outliers. c Scene prediction accuracies compared between trials with high and low state-confidence levels estimated by the HMM. The prediction accuracy was significantly higher when the HMM estimated the confidence level is high compared to low (one-sided Wilcoxon signed-rank test, ***: p < 0.001). The dashed line indicates chance level. d Decoding accuracy of the HMM estimated state confidence. The decoding accuracy was evaluated using LOSO CV (one-sided Wilcoxon signed-rank test, ***: p < 0.001). The dashed line indicates the chance level. Here we plotted the result for the 6th decoding period representatively; the results of our time-series decoding analysis are shown in Supplementary Fig. 7c. e Time-series decoding accuracies of scene prediction for different state-confidence levels. Each decoder for each period was trained with the trials of high (or low) state confidence estimated by the HMM, and the accuracy was evaluated by LOGO CV. The solid lines represent the median, shaded areas indicate the range between the upper and lower quartiles, and the dotted lines indicate the range of 1.5 × IQR. Cross-markers indicate the outliers. Significance was tested using a one-sided Wilcoxon signed-rank test (unfilled circle: p < 0.05, unfilled square: p < 0.01, unfilled diamond: p < 0.001) compared to the chance level (dashed line). The colors of the horizontal line below the plots reflect the significant differences between the two categories of the trials in each decoding period (one-sided Wilcoxon rank-sum test, light gray: p < 0.05, dim gray: p < 0.01).