Introduction

The neural mechanisms underlying navigation strategies have received considerable interest from both experimental and theoretical research^1–4. However, the fact that multiple strategies might be used within the same environment adds another layer of complexity to this problem^5–9. How navigation strategies are selected is an intriguing question that needs to take into consideration several factors. One important factor is the nature and availability of spatial cues, which can largely vary between environments. Wide open fields such as deserts or city plazas are believed to promote vector-based strategies that rely on distal landmarks and the elaboration of a cognitive map^{3, 10–13}, whereas intricate networks of channels/streets should privilege response-egocentric or sequential-egocentric strategies that rely on local cues and memory of left/right turns^14–16. Another factor is the level of familiarity with the environment. Exploratory behaviors are expected to predominate in novel environments, whereas vector-based navigation and egocentric strategies should progressively increase as animals become more familiar with landmark layouts and routes to goals, respectively. Additionally, the motivational factors driving animal motion in the environment, such as reward-seeking, threat-avoidance, and uncertainty resolving, can also affect navigation strategy^{5, 17, 18}.

The Barnes maze provides a suitable platform for examining the repertoire of strategies and the shift in strategy that accompanies learning of the environment¹⁹. In this task, rodents explore a circular arena until they find an escape box located under one of several holes distributed along the arena periphery. Simple criteria are commonly used to classify the strategy employed in each trial, which are largely based on the number of visited holes and the assumption that the same strategy is maintained across an entire trial^20–23. However, it is largely unclear whether these criteria are valid and accurately capture the full range of strategies used.

To address these issues, we developed an automated variant of the Barnes maze and performed learning, probe test, and reversal learning experiments commonly carried out in Barnes mazes. Furthermore, we developed new analytical tools to characterize the statistical patterns of vestibule visits; generated vestibule sequences via stochastic processes intuitively compatible with random, serial, and spatial strategies; and demonstrate that a combination of these stochastic processes can largely account for vestibule visit patterns. These amount to a new method for tracking navigation strategies based on the relative weights of stochastic processes, which we used to demonstrate a shift from random to spatial and serial strategies across days and a strategy turn-over period of approximately 6 vestibule visits.

Results

Automated maze

We developed a fully automated maze apparatus allowing to randomize the start and goal positions of individual trials without the need for experimenter intervention (Fig. 1, Supplementary Fig. 1). Our maze is similar to the Barnes maze in that mice explore an open arena and are required to navigate to a goal location chosen among an array of 24 doors evenly spaced along the periphery of the arena. However, unlike typical Barnes mazes where orienting cues are external to the maze and animals start at the center of the maze, our maze features two luminous objects inside the arena as orienting cues, and start positions are located at the periphery of the arena.

Figure 1

Behavioral protocol

For a first block of 15 days (acquisition phase), mice performed 10 trials per day. In each trial, they entered the maze from a randomly chosen start position and navigated to the goal position to consume a water reward. After the 15 days, a probe test was performed, during which the goal door remained closed and mice were allowed to explore the arena for 2 minutes. Next, for another block of 4 days (reversal phase), mice performed the same task as in the initial block of 15 days, but the goal position was changed. We recorded behavioral data from 10 male and 10 female mice (C57BL6, 7-8 weeks old), and since very similar patterns were observed for the two groups, we pooled and analyzed the data together in the main figures (male and female groups are analyzed separately in supplementary figures).

Segmentation of spatial trajectories

To analyze the spatial behavior of mice, we decomposed their trajectories into segments that spanned from one vestibule to another, as mice tended to explore the arena and visit multiple vestibules before reaching the goal. To accomplish this, we first detected vestibule visits, and then measured the number of door-intervals each segment spanned, considering the direction of travel associated with the shortest distance. A segment could span from 1 to 12 door-intervals in either the clockwise or counterclockwise direction (Fig. 2a). Positive door-interval values represented clockwise directions, while negative values represented counterclockwise directions.

Figure 2

Statistical characteristics of spatial trajectories across days

We analyzed the statistical characteristics of spatial trajectories across days (Fig. 2 and Supplementary Fig. 2). We found that the path length of individual trials decreased across days, with the largest drop observed between day 1 and day 2 (Supplementary Fig. 3a), consistent with previous reports in the Barnes maze^{21, 24, 25}. However, the path length of individual segments did not show substantial change across days when comparing segments of equivalent spans and directions (Fig. 2b), suggesting that the overall decrease in trial path length was more associated with a decrease in the number of vestibule visits rather than an alteration of running paths. Next, we calculated the trial length (number of segments per trial). We found that the average trial length decreased across days, with the largest drop observed between day 1 and day 2 (Supplementary Fig. 3b). The distribution of trial lengths showed a peak for short trials followed by a progressive decline toward longer trials (Fig. 2c). The proportion of short trials became higher across days, at the expense of longer trials, consistent with the average decrease in trial length.

To determine whether mice relied on allocentric spatial information to learn the goal location, we analyzed the proportion of visits to each vestibule, expecting to see an increasing concentration of visits near the goal location over time. Although no preference was observed on day 1, mice progressively showed a preference for the goal vicinity in subsequent days, supporting their use of allocentric cues to locate the goal (Fig. 2d). To further validate this finding, we measured the time mice spent in each quadrant of the arena and in each vestibule during the probe test. The percentage of time spent in the goal-containing southwest (SW) quadrant was significantly higher than chance (p = 2.54×10⁻⁴, two-sided Wilcoxon sign-rank test) and other quadrants (Supplementary Fig. 4a, p <0.001, two-tailed paired t-test), and the percentage of time spent in the goal vestibule was significantly higher than chance (p = 0.009, two-sided Wilcoxon sign-rank test) and other vestibules (Supplementary Fig. 4b, p <0.05, two-tailed paired t-test), confirming that mice relied on allocentric cues to learn the goal location. In the Barnes maze, mice occasionally exhibit a search strategy called serial, where they visit consecutive holes in series. To determine the extent of this behavior, we calculated the length of segments (the number of door-intervals they spanned; Fig. 2a). On day 1, there was a high proportion of short segments (1 to 2 door-intervals), consistent with the use of serial strategy (Fig. 2e, F_2,108 = 3.66, p = 0.03, two-way ANOVA). This proportion increased over subsequent days, indicating an increased reliance on serial strategy (14.09 ± 5.81 %, 17.50 ± 5.39 %, 18.96 ± 6.1 % for day 1, day 2, and day 6-15, respectively; p = 0.001 for day 1 vs day 2, and p = 0.002 for day 1 vs day 6 - 15, two-tailed paired t-test).

Interestingly, we found that the peak of the segment length distribution was more than three times (3.56 ± 1.13 time, across days) larger for the clockwise than the counterclockwise direction, indicating that the serial strategy was mostly used in the clockwise direction. This was likely due to the vestibules’ orientation, which determined mice body orientation as they left a vestibule, matching the clockwise direction. When we switched the orientation of the vestibules, we observed a reversal of the animals’ serial visit direction from clockwise to counterclockwise (Supplementary Fig. 5).

Additionally, we computed the length of serial bouts, defined as the number of consecutive one-door-interval-long segments, and found that the proportion of serial bouts sharply decayed as a function of serial bout length, with 1-segment-long bouts accounting for 82.62 ± 13.16 % of the bouts on day 1 (Fig. 2f). The proportion of longer bouts was increased across days, consistent with an increased usage of serial strategy (17.37 ± 13.16%, 29.34 ± 17.14%, 32.68 ± 11.38% for day 1, day2, and day 6-15, respectively).

Statistical characteristics of spatial trajectories across trials

To determine if the pattern of vestibule visits changed across individual trials within a session, we pulled together the data from day 6 to 15 and performed the same analyses as in Figure 2 and Supplementary Figure 2 for individual trials (Fig. 3 and Supplementary Fig. 6). We found that changes across trials were not observed for the distribution of trial lengths and vestibule visits (Fig. 3b and 3c), but the distributions of segment length and serial bout length showed an evolution across trials that resembled the evolution across days. Specifically, the first trial showed profiles similar to day 1, and later trials showed profiles similar to later days (Fig. 3d and 3e). Hence, a similar evolution of serial strategy was observed across days and trials.

Figure 3

Stochastic processes for random, spatial and serial strategies

To investigate if the observed pattern of vestibule visits could be explained by random, spatial, and serial strategies, we implemented three distinct stochastic processes (Fig. 4a), performed simulations using same number of mice and trials and the same start positions as in the experiments, and performed the same analyses as Figure 2 on the simulated data (Fig. 4b). Each strategy determined which vestibules would be visited next and was recursively run until the goal was reached.

Figure 4

For the random strategy, the identity of the next vestibule was randomly picked based on a uniform probability distribution, which produced a relatively uniform distribution of vestibule visits, similar to day 1 of the experiment. For the spatial strategy, the identity of the next vestibule was randomly picked based on a probability distribution that exponentially decayed as a function of goal distance. This strategy reproduced the concentration of vestibule visits near the goal observed in days 2 to 15, and partially contributed to the peaks in the distribution of segment lengths. For the serial strategy, the identity of the next vestibule was determined by incrementing the current vestibule number by a step, with the sign of the step randomly picked based on an 80% chance of being positive (clockwise direction) and the size of the step obtained by adding a normal random jitter to the value of 1.2 for the clockwise direction and 2 for the anti-clockwise direction. This strategy reproduced the distribution of segment lengths.

Models Combining Random, Spatial, and Serial Stochastic Processes

None of the strategies by themselves could reproduce all of the experimental distributions shown in Figure 2. We therefore implemented models that used a combination of the three strategies, where strategies were stochastically selected according to a particular distribution of probabilities (Supplementary Fig. 7). To find the optimal probability distribution, we ran simulations for all possible combinations of probabilities, with each probability ranging from 0 to 100 percent with an increment of 2. For each combination, we performed the same analyses as Figure 2 and compared the results with experimental data by computing mean square errors. The optimal probability distribution was the one generating the lowest mean square error (Fig. 5a). To track random, serial, and spatial strategies over the 15 days, we repeated all simulations and error calculations for each day, using the same sequences of start positions as in the experiments.

Figure 5

We first tested two models: one where mice chose one of the three strategies at the start of the trials and used the same strategy throughout individual trials, and another where mice chose a strategy each time they left a vestibule (i.e. for each segment). Both models reproduced the general features of experimental data across days. However, some distributions were systematically overestimated by one model and systematically underestimated by the other, suggesting that a model with an intermediate rule for strategy selection might be optimal. Therefore, we implemented a third model in which mice chose one of the strategies at the beginning of the trials and then each time they completed a particular number (N) of segments. To find the optimal value of N, we tested a range of values from 1 to 10 and selected the one generating the lowest mean square error. This third model best reproduced the experimental data (Fig. 5a and Supplementary Fig. 8a), generating the lowest mean square errors among the three models (Fig. 5b). The number N was larger than 1 and smaller than 10 for all days and had an average value of 6.29 ± 0.64 (Fig. 5c). These results demonstrate that the combination of the three stochastic processes could account for the patterns of vestibule visits.

Evolution of search strategy across days

Using the probabilities obtained from the third model, we examined how the proportion of random, spatial and serial strategies evolved across days (Fig. 5d and Supplementary Fig. 8b). The spatial strategy was increased across days, from 13.4 ± 4.16 % on day 1 to 53 ± 8.34 % on day 15 (p = 1.86 x 10⁻⁶, two-tailed paired t-test), largely at the expense of the random strategy, which decreased from 58.2 ± 3.58% on day 1 to 3.8 ± 4.16 % on day 15 (p = 6.94 x 10⁻¹¹, two-tailed paired t-test), while the serial strategy increased mostly from day 1 to day 2 (28.4 ± 3.1 % on day 1 versus 44.6 ± 3.27 % on day 2, p = 7.62 x 10⁻⁷, two-tailed paired t-test).

To compare our results with previous studies, we also assessed the proportions of random, spatial, and serial strategies using the methods used in those studies (Supplementary Fig. 8c and d). Specifically, we classified a trial as using a spatial strategy if the number of vestibule visits before reaching the goal was less than 3, a serial strategy if the goal was reached via a serial bout of at least 3 segments, and a random strategy if the number of vestibule visits exceeded 3 and no serial bout led to the goal. Consistent with our model results, we found that the spatial and serial strategies increased across days, while the random strategy decreased. However, the proportions of the random strategy remained relatively high across all days (∼ 46% on day 15), in contrast to the model results.

Discussion

In this study, we developed an automated variant of the Barnes maze. In general, automated behavioral apparatuses reduce experimenter interference and ensure consistency across time and individuals^26–29. Our maze was similar to the Barnes maze but had distinct features such as vestibules in front of the doors, which made detection of vestibule visits relatively easy, and the use of two luminous objects in the arena for allocentric information. Despite these differences, we found that learning rates and spatial strategies were comparable to those reported in the Barnes maze^{21, 24, 25}. Furthermore, we found that mice had a preferred direction of rotation during the serial strategy, which was determined by the physical configuration of vestibules. Directional preference was observed in several other studies^30–32, which may have also originated from environmental factors.

Our analyses of vestibule visits revealed that mice employed random, spatial, and serial search strategies. While these strategies have been previously considered^20–23, this is the first time that compatible stochastic processes are implemented and shown to complementarily account for exploration patterns. The spatial strategy likely used a form of allocentric navigation given the distance of the two luminous landmarks from the goal, and may be divided in two phases: the learning of the environment and goal location, and the navigation to the goal. The learning of the environment and goal location might be supported by the formation of a cognitive map in the hippocampal formation^{10, 33–35}, achieved through competitive learning^36–38, which provides the substrate for binding place and reward information^39–41. The navigation to the goal may rely on vector navigation, where goal vectors^{42, 43} are implemented via the comparison of grid cell codes for goal and current positions^{11, 44, 45}, or on topological navigation, where the trajectory to the goal emerges from signal propagation in the place cell network^46–48, and may be assisted by lookahead mechanisms such as theta sweeps and/or ripple-associated replays^49–51. The mechanism behind the serial strategy is less clear. Although the repetition of left turns is reminiscent of an egocentric response, the serial strategy was observed from day 1, indicating that it did not require learning, unlike the egocentric response navigation supported by the dorsolateral striatum^{5, 52, 53}. The serial strategy might instead share mechanisms with thigmotaxis, an innate behavior that is enhanced by anxiety and characterized by animals staying close to walls as they explore an open environment^{17, 52, 54}. However, in our experimental paradigm, this behavior may have been promoted by the vestibule-reward association rather than anxiety. Finally, the random strategy might be related to exploratory behaviors and engage an ensemble of brain regions important for sensory integration, decision-making and reinforcement learning such as the hippocampus, prefrontal cortex, and basal ganglia^{9, 55–62}.

A shift in the proportion of random, spatial and serial strategies was observed across days. The spatial strategy progressively increased, mostly at the expense of the random strategy, while the serial strategy showed a sudden increase from day 1 to day 2. One possible explanation for this shift is that the mice continuously evaluated the benefits of each strategy using a reinforcement learning process that integrate rewards over several trials. As the relationship between the goal and allocentric cues is learned, the spatial strategy becomes increasingly beneficial and is therefore more frequently used. Consistent with this hypothesis, the development of the spatial strategy approximately matched the development of spatial maps in the hippocampus³⁸ and the growth pattern of hippocampal feedforward inhibitory connectivity⁶³, which both showed progressive increases that reached plateaus after a week. In contrast, the serial strategy, which does not require learning, can be employed from day 1 and is already reinforced on day 2 as it generates the highest reward.

Maintaining the same strategy across several trials is particularly useful for successfully reaching the goal via the serial strategy, and might rely on mPFC working memory and strategy selection functions^64–66. Our model and experimental data were optimally matched when the same strategies were maintained for about six segments. Most trials were less than 6-segment-long (∼60% and ∼66% of trials on day 1 and 15), and therefore involved the use of the same strategy throughout the whole trial. However, a substantial number of trials were longer and therefore involved switching between strategies. This suggests that commonly used methods for categorizing navigation strategy on a trial-by-trial basis may not accurately capture the full repertoire of strategies employed.

In conclusion, we have developed a fully automated variant of the Barnes maze and some analytic tools allowing to track the repertoire of navigation strategies over time. Furthermore, we have demonstrated that a set of stochastic processes compatible with random, spatial and serial strategies could largely account for vestibule visit patterns. Our tools might be combined in the future with optogenetic and/or pharmacogenetic to investigate the neural mechanisms underlying strategy selection, whereas future work in silico might investigate neuronal network models generating exploration patterns equivalent to the stochastic processes.

Materials and Methods

Maze description

The maze apparatus consists of an aluminum frame, a central arena and two home boxes that can move around the arena (Fig. 1a and Supplementary Fig. 1). The floor of the arena (a white acrylic disc 1-cm-thick and 95-cm wide) is supported by a mounted bearing (UC205-16, UCFT205-16) fixed to a central rod (2.54 cm-diameter, 23 cm-high, stainless steel) and by 4 wheels at a 14.5-cm distance from the periphery. One of the wheels is motorized (NEMA 17HS4401 Bipolar Stepper Motor) to control the rotation of the platform. The angular position of the floor is calculated from the motor increment signal and an LED-photodetector couple detecting the passages of 24 beam-breakers (2×0.5×20 mm plastic pieces) fixed under the floor at positions matching the arena doors.

The arena is enclosed by a 16-cm-high peripheral wall made of white acrylic (thickness: 0.3 cm) with 24 evenly distributed entrances 8-cm apart from each other. Each entrance consisted of a 5-cm hole with a flapping door made from a 3D printer (ABS, Stratasys/F120).

The entrance doors are normally closed, but following the alignment with a home box, are rotated into open positions throughout the opening of the box door. The arena is partitioned by 7-cm-high internal walls made of styrofoam blocks, on which a transparent-tainted acrylic cover (50-mm-thick) is laid to prevent mice from leaving the maze and seeing the room cues while allowing mouse tracking with infrared light.

The home boxes are made with a 3D printer (ABS, Stratasys/F120) and have motorized rotating doors, allowing mouse access to the maze, and an opening on the top, allowing experimenter access to the mice. Upon entering the box, mice have to turn ∼180 degree around an internal wall to reach a water dispenser, ensuring that no body parts (such as the tail) remain near the door when it closes. An LED-photodetector couple triggers the water reward and door closure upon the mouse reaching the water dispenser.

The home boxes are coupled to the central axle of the maze through an aluminum arm and a mounted bearing. A motorized wheel fixed under the home boxes allows the home boxes to roll on an aluminum platform along the circumference of the arena. The angular position of the box is calculated from the motor increment signal and an LED-photodetector couple detecting the passages of 24 beam-breakers (2×0.5×20mm plastic pieces) fixed under the aluminum platform at positions matching the arena doors.

To control the maze, a microcontroller (Arduino mega 2560) is used. The home box electronics are connected to the microcontroller through a long ribbon cable, which is placed inside a flexible plastic accordion pipe (148 cm-long, 25 mm-diameter) to prevent cable entanglement during the rotation of the boxes.

Maze environment

For the spatial navigation task, the arena was accessed via angled vestibules, which limited the visibility of doors from within the arena (Supplementary Fig. 1). Two luminous landmarks were placed on the arena, consisting of small (WxLxH, 5×5×5 cm) and large (10×15×5 cm) semi-transparent plastic boxes containing yellow and blue LEDs, respectively.

Experimental subjects

All experiments were conducted in accordance with institutional regulations (Institutional Animal Care and Use Committee of the Korea Institute of Science and Technology) and conformed to the Guide for the Care and Use of Laboratory Animals (NRC 2011). Ten female and ten male (C57BL6, 7-8 weeks old) were used for the experiment. The mice were housed in a vivarium with 5 mice per cage (males and females were kept separate) and maintained under a 12-hour light/dark cycle. All experiments were carried out during the light cycle.

Task and maze operation

The task involved navigating from a random start position to a fixed goal position (Fig. 1b). At the beginning of each session, a mouse was placed in one of the home boxes. The start position was determined by rotating the arena floor until the box was aligned with the desired start position, while the goal position was set by moving the other box until it was aligned with the desired goal position. The doors of the boxes opened to let the mouse navigate the maze and then closed when the mouse reached the lick port of the other box. For most trials, the box containing the mouse remained still while the other box and arena floor were moved. However, both boxes were moved when the rotation angle of one of the boxes was going to exceed 180 degrees to limit the twisting of electric cables.

Mice performed a total of 10 trials per session, for 15 sessions (1 session per day), with a distinct sequence of random start positions for each session. The same sequence of start positions was used for all mice. All these steps were carried out automatically using a program uploaded to a microcontroller using Arduino IDE 1.8.19.

Probe test

Probe tests were carried out 3 hours after the task, on day 15. The mice were placed in the arena and allowed to explore freely for 2 minutes, with all arena doors closed. The maze area was divided into 24 equal zones, with the goal position at the center of one of the zones. The time spent in each zone and quadrant was measured using Ethovision (Noldus, Spink, & Tegelenbosch, 2001).

Trajectory tracking

Animal behavior was recorded using an infrared video camera (Basler acA1300-60gm, 25 frame/sec) and infrared lights (AP-XM722-WAB-U4), and synchronized with task signals from the microcontroller using the software Ethovision (Noldus, Spink, & Tegelenbosch, 2001). Mouse position tracking and further analyses were implemented in MATLAB (MathWorks Inc., MA, USA). Video frames from a period without the mouse (the period between stopping the rotations of the boxes/arena floor and the opening of the door) were used to compute an average baseline frame that was subtracted from subsequent video frames with the mouse on the arena. For each frame, pixels delineating the mouse area were detected (using a threshold) within a circular region (radius: 7 cm) centered initially on the start position and then on the previous mouse position. The mouse area was longitudinally divided into 4 equal segments, and the center of mass xy-coordinates were computed for each segment, allowing a tracking of mouse position and body orientation. For each trial, the xy-coordinates were rotated to align the maze environment and goal position across trials.

Trajectory segmentation

On each trial, mice often visited several vestibules before reaching the goal box. We divided each trial’s trajectory into two distinct components: vestibule visits and running segments between vestibules. We defined vestibule entrances and exits as the points where the mice crossed a radial distance threshold of 47.5 cm from the center of the arena.

Path length

Mouse displacement was computed for each time step based on Euclidean distance and integrated over the entire trajectory to obtain the path length. We measured path length for both individual trials and individual segments, defined as the path between two vestibules.

Trial length

The trial length was defined as the number of segments in a trial.

Segment length and direction

The segment length was defined as the number of doors it spanned, considering the travel direction associated with the smallest number. Each segment could span from 1 to 12 doors, either in a clockwise or counterclockwise direction (see Fig. 2a). A positive value of segment length indicated a clockwise direction, while a negative value indicated a counterclockwise direction.

Serial bout length

For each trial, we calculated the length of segments and identified consecutive 1-door-long segments, which we defined as serial bouts. We measured the length of serial bouts as the number of segments in the group, with a minimum length of 1 segment corresponding to single, isolated 1-door-long segments.

Stochastic process for random strategy

To model the pattern of vestibule visits, stochastic processes were implemented to choose which of the vestibules was visited next (Fig. 4a). For the random strategy, the identity of the next vestibule was randomly drawn from the uniform distribution in the range [1 24], using MATLAB function randi to generate uniformly distributed pseudorandom integers in the range [1 24].

Stochastic process for spatial strategy

For the spatial strategy, the identity of the next vestibule was randomly drawn from a distribution P that exponentially decayed as a function of goal distance (number of door-intervals between the vestibule and the goal). The distribution P was characterized by two symmetrical negative exponentials exp(-x/tau) that decayed as a function of leftward/rightward goal distance. An integer was randomly drawn from the uniform distribution in the range [1 sum_of_the_distribution_P]. The goal distance was determined by the position x for which the cumulative sum of the distribution P reached the value of the integer. The decay rate (tau) of the exponentials was heuristically set to 2 through comparison of experimental and simulated data (Fig. 2 and 4). Importantly, we assumed that the next vestibule could not be the same as the current vestibule for this goal oriented strategy. Therefore the stochastic process was run again if the current vestibule was selected as the next vestibule.

Stochastic process for serial strategy

For the serial strategy, the identity of the next vestibule was determined by incrementing the current vestibule number by a stochastically determined step.

The sign of the step was randomly drawn, with an 80% (20%) probability of being positive (negative), which indicated a step in the clockwise (counterclockwise) direction. Specifically, an integer x was randomly drawn from the uniform distribution in the range [0 100]. If x < 80, the sign was positive, otherwise it was negative.

The size of the step was obtained by adding a random jitter to an offset value (1.2 and −2 for clockwise and counterclockwise directions, respectively). The jitter was randomly drawn from the normal distribution, using MATLAB function randn to generate normally distributed pseudorandom numbers, and was multiplied by a standard deviation value (1.2 and 1.5 for clockwise and counterclockwise directions, respectively).

The values of offsets, standard deviations, and sign probabilities were heuristically determined through comparison of experimental and simulated data (Fig. 2 and 4). Specifically, for the two peaks observed in the segment length distribution, the offsets determined the positions of the peaks, the standard deviation determined the width of the peaks, and the sign probabilities determined the ratio of the height of the peaks.

As for the spatial process, we assumed that the next vestibule could not be the same as the current vestibule, so the stochastic process was rerun if the current vestibule was selected as the next vestibule.

Models combining the 3 strategies

We implemented models that combine the three strategies (random, spatial and serial) to simulate the pattern of vestibule visits (Supplementary Fig. 7). Each strategy was assigned a specific probability (P_random, P_spatial, P_serial) for a given simulation. To select the strategy, an integer x was randomly drawn from the uniform distribution in the range [0 100]. The strategy was random if x < P_random; it was serial if x > P_random + P_spatial; otherwise, the strategy was spatial.

We simulated the models for the same number of mice and trials as in the experiments, and used the same start positions as in the experiments.

The first model assumed that mice used the same strategy throughout a trial. At the beginning of each trial, a strategy was randomly selected and the stochastic process associated with the selected strategy was recursively used to draw the next vestibules until the goal was reached.

The second model assumed that mice selected a strategy each time they left a vestibule. For each trial, both the strategy selection and the stochastic draw of the next vestibule were recursively executed until the goal was reached.

The third model had the intermediate assumption that mice used the same strategy throughout a specific number n of consecutive vestibules. At the beginning of a trial, a strategy was randomly selected and used to draw the next n vestibules. Then, the strategy was randomly selected again and used to draw the next n vestibules, and so on, until the goal was reached.

Estimation of the probability of the strategies

To determine the optimal probability values for each strategy (P_random, P_spatial, P_serial), we conducted a comprehensive simulation study, exploring a wide range of probability combinations. Specifically, we tested values of P_random, P_spatial, and P_serial ranging from 0 to 100 (percent) with an increment of 2, while ensuring that their sum was always equal to 100. For each combination of probabilities, we used the same analytical methods as in the experimental analysis and computed the mean square error between the two distributions. We selected the optimal probability values as those that produced the lowest mean square error between the simulated and experimental distributions.

To determine P_random, P_spatial and P_serial for each of the 15 days, all simulations and error calculations were repeated for each day, using the same number of mice, trials and the same start positions as in the experiments.

Author contribution

JYL and SR designed the project and built the maze with participation of DJ. JYL performed the experiments. JYL and SR performed analysis and wrote the manuscript.

Acknowledgements

This work was supported by the Korea Institute of Science and Technology Institutional Program (Project # 2E32211) and the National Research Foundation of Korea (NRF Grant # 2N69930).

Additional information

Competing Interests: The authors declare that they have no competing interests.

Data availability

The data that were collected for this study are available upon reasonable request.

Code availability

All the codes for data analysis and modeling are available upon reasonable request.