Presentation Information
[P3-03]A Deep Reinforcement Learning Approach to Modeling Rat Behavior in Peak Interval Procedure
*S. Ruiz de Aguirre1, Gloria Ochoa-Zendejas2, Jonathan Buriticá2 (1. Independent (Mexico), 2. Lab. of Cognition and Comparative Learning, Univ. of Guadalajara-CEIC, Guadalajara (Mexico))
Keywords:
Timing,Neural Networks,Peak Interval Procedure,Computational Modeling,Animal Behavior Simulation
Time estimation is an important concept of adaptive behavior. Most studies focus on utilizing Peak Interval Procedures or Time Bisection Tasks, commonly done utilizing animal models. While animal models may accurately represent biological features, they come with ethical and economical caveats, also being time-consuming. In this study, we intend to generate a computational model to simulate the behavior of rats in Peak Interval Procedures with the objective of providing a replicable and low-cost alternative to running the same experiment on actual animals.
The proposed process utilizes deep reinforcement learning to generate agents that replicate previous empirical data from real rats in Peak Interval Procedures, aiming to achieve a similar Gaussian-like distribution with a peak centered around a 30-second target interval. Agents will be trained using reinforced fixed intervals, and evaluated after each training epoch in fifteen non-reinforced Peak Interval Procedure trials, until achieving results similar to the empirical data; at that point, model weights will be stored. The training process will take into account the configuration of the operant box and penalizations for energy expense upon any action not providing reinforcement.
We expect the model to replicate the characteristic peak in responding around the target interval and to generalize across different durations with adequate training. Beyond its theoretical relevance, this solution may offer an ethical and economic advantage: reducing the number of animals used in experimental settings. This work represents a step toward integrating computational intelligence with animal models in behavioral analysis for timing.
The proposed process utilizes deep reinforcement learning to generate agents that replicate previous empirical data from real rats in Peak Interval Procedures, aiming to achieve a similar Gaussian-like distribution with a peak centered around a 30-second target interval. Agents will be trained using reinforced fixed intervals, and evaluated after each training epoch in fifteen non-reinforced Peak Interval Procedure trials, until achieving results similar to the empirical data; at that point, model weights will be stored. The training process will take into account the configuration of the operant box and penalizations for energy expense upon any action not providing reinforcement.
We expect the model to replicate the characteristic peak in responding around the target interval and to generalize across different durations with adequate training. Beyond its theoretical relevance, this solution may offer an ethical and economic advantage: reducing the number of animals used in experimental settings. This work represents a step toward integrating computational intelligence with animal models in behavioral analysis for timing.