2026-09-21T15:36:26Zhttps://repository.rice.edu/server/oai/request

oai:repository.rice.edu:1911/1186732025-09-16T20:45:55Zcom_1911_8299col_1911_13110

Unhelkar, Vaibhav Pitkow, Xaq Olivos Castillo, Itzel Coral 2025-09-11T16:01:14Z 2025-09-11T16:01:14Z 2025-08 2025-08-13 August 2025 2025-09-11T16:01:14Z https://hdl.handle.net/1911/118673 Today's machines rely on resource-intensive algorithms and specialized hardware to produce intelligent sensorimotor behavior. In contrast, the brain can transform noisy stimuli into effective actions that solve a wide range of tasks using limited experience, operating with modest processing capacity, and consuming less energy than a lightbulb. Neuroscientific research suggests that to achieve this remarkable performance, the brain not only reasons about external factors but also monitors and regulates its internal processes. Incorporating this meta-cognitive ability into artificial systems is attractive, as it may assist agents in better aligning their computational efforts with resource constraints, task requirements, and environment structure. However, the practical implementation of meta-cognition remains an exciting open challenge, especially in uncertain environments. To help bridge this gap, we propose a novel approach to stochastic control that enables the regulation of inference—an internal process through which agents transform noisy observations into beliefs that guide their behavior. We apply our framework to quantitatively examine how meta-cognitive agents solve environments with linear dynamics, Gaussian sources of noise, and quadratic cost functions (LQG environments). Our study reveals that meta-cognition strongly intertwines inference and control dynamics. This coupling reshapes the convex optimization landscape of classic LQG control and leads to intriguing phase transitions in what is worth being optimally inferred. Based on environmental reliability, task demand, and resource availability, the strategies of meta-cognitive agents switch from a costly mechanism that relies on Bayes-optimal inference to multiple combinations of information-efficient inference and error-aware control. Our findings generalize efficient coding ideas in neuroscience, extend the principle of minimal intervention in control, and offer valuable insights that, combined with advancements in low-power hardware, could propel the development of a new wave of artificial intelligent systems capable of matching the outstanding resource efficiency of their biological counterparts. application/pdf eng Copyright is held by the author, unless otherwise indicated. Permission to reuse, publish, or reproduce the work beyond the bounds of fair use or other exemptions to copyright law must be obtained from the copyright holder. Meta-cognition Bounded Rationality Approximate Inference POMDP Efficient Control Theoretical Principles for Information-Efficient Reasoning under Uncertainty Thesis Text Computer Science Engineering Rice University Doctoral Doctor of Philosophy open.access