Probabilistic model of rat movements considering inertia. Training and test datasets collected while manually operating a VCR.


  • The model is focused on predicting the behavior of healthy rats without any treatments.
  • Tells about the preference of rats to stay close to vertical surfaces.
  • Talks about “unrealistic assumption” from Salum, Morato, & Roque-da-Silva (2000) that rats make exactly one movement every two seconds.
  • Observed that real rats rarely makes more than one movement per second.
  • Could not reproduce results from Salum et al. (2000). Possibly due to their small sample size.
  • Used Cartesian coordinate system to represent the maze. Three squares for each of the arms, an extra square for the center, totaling thirteen squares.
  • Provides Matlab source code for the model.
  • In contrary to Salum et al. (2000), does not include squares outside of the maze, but has a parameter to represent if the rat is acting in rearing or risk assessment behavior.
  • Started with simulations considering equal probabilities for going east (0, 0.2), west [0.2, 0.4), north [0.4, 0.6), south [0.6, 0.8), and standing still [0.8, 1). Tested the model multiple times and adjusted values to reproduce theories mentioned by Salum et al. (2000). Does this to start a better model grounded on existing behavioral theories.
  • Probability model considers moving outward or inward in open or closed arm, or to the center.
  • Based on 34 trials with real rats. Watched videos, manually paused the VCR, and took note of the position.
  • Interviewed researchers who said that time decay parameter was not necessary. Five minutes doesn’t make a lot of difference since times in each of the arms will be averaged out in the end.
  • “Final Inertia Model” comes with probabilities addressing expectations that rats will keep moving if was moving before, unless when going to open arms or to the center. In this case, the probability is smaller.
  • Used t-tests, as Salum et al. (2000), to compare times of real and virtual rats. Initial test raised concerns about the effectiveness of the model. Changed one of the distribution values and results got better. Not so much when looking to entries in open arms, but, averaged out, turned out acceptable.
  • While trained with 39 rats, the model was tested with other 69 real rats. Compared with 69 predictions from virtual rats.
  • In the end, reconsidered the importance of Montgomery’s observation that rat behavior will change over time.


  • Since this model has one second as the unit of time, it would be interesting to decrease even more, going to the limit of zero seconds. In practice, this would be considered the same as a continuous time.
  • Again, the sample size seems to be quite small (n=34). Representative models may require more rats to train over.
  • Most of the available rats were used for testing, not for training, suggesting the possibility of underfitting.
  • The author stated that there’s no function to be optimized. I believe parameters could be optimized using Supervised Machine Learning while comparing to data from real rats.


Salum, C., Morato, S., & Roque-da-Silva, A. C. (2000). Anxiety-like behavior in rats: a computational model. Neural Networks, 13(1), 21–29.