Skip to main content

Bounded Rationality in Sensorimotor Coordination

Periodic Reporting for period 3 - BRISC (Bounded Rationality in Sensorimotor Coordination)

Reporting period: 2019-10-01 to 2021-03-31

Machines can repeat the same movement a million times with relentless precision, while humans struggle even after years of practise to hit the bull’s-eye twice in a row. Compared to humans, machines seem to have superior resources ranging from more reliable hardware, better signal-to-noise ratios in their signal processing, less sensorimotor delays, and higher computation speed. Yet, despite their successful application in well-structured and predictable environments, machines are still blatantly outperformed by humans in unstructured and uncertain environments that require flexible behavior through real-time sensorimotor interactions like playing football or navigating a disaster zone. The goal of this project is to elucidate the effect of limited computational resources in biological sensorimotor processing and to investigate how control strategies that take such limitations into account might cope better in real-world scenarios with model uncertainty and computational constraints. In order to achieve this goal the project aims to establish a bounded rational framework for sensorimotor processing that allows unifying action and perception within the same formalism and considers information-processing costs as the fundamental currency. Establishing this novel framework requires drawing out theoretical predictions and gathering experimental evidence along three complementary objectives that explore predictions of the same basic concepts in three different but interdependent areas in single-agent, multi-agent and hierarchical motor control. In particular, the objectives are to gather experimental evidence of bounded rationality in human motor control and compare to different bounded rationality models, to investigate the effects of bounded rationality on sensorimotor interactions between multiple actors from the point of view of bounded rational game theory, and to instigate the hypothesis that bounded rationality of individual information-processing nodes naturally leads to the formation of hierarchical control structures and the development of different levels of abstraction from the sensorimotor stream. Understanding the principles underlying the organization of systems composed of bounded rational agents and the pertinent organization of abstractions provide a decisive step towards understanding autonomous intelligent systems that develop their own concepts in interaction with the world.
The main results in this first reporting period fall into three broad categories: (i) theoretical advances in conceptualizing bounded rational decision-making, (ii) technical applications to sensorimotor learning in robotics, and (iii) biological applications to sensorimotor learning in humans.

(i) Theoretical advances: We have developed a theoretical framework that justifies conceptualizing bounded rational agents by an information-utility trade-off through the more basic concept of elementary computations, simple transformations between probability distributions that reduce uncertainty. Moreover, systems that can be described by an information-utility trade-off are amenable to a statistical description analogous to statistical thermodynamics, where we could show that bounded rational decision-making processes can also be studied with the tools of non-equilibrium physics. Finally, we could show how the information-utility description can be used to study and to optimally design the organization of systems of multiple bounded rational agents with unidirectional information flow.

(ii) Technical applications: We have studied how the principles of bounded rational decision-making can be applied to simple robotic systems, in particular how abstractions can be formed as a consequence of limited information. We have also proposed artificial neural network models that can be used in conjunction with planning modules to implement bounded rational decision-making.

(iii) Biological applications: We have devised simple sensorimotor and stimulus identification tasks with humans and demonstrate how the bounded rationality framework can be used to quantify behavior, in particular whether behavior is close to bounded optimal, and how to use this measure to instigate investigations for sources of inefficiency.
Research on sensorimotor processing has largely relied on Bayes-optimal models for studying simplified and well-controlled laboratory tasks. However, Bayes-optimal models of sensorimotor processing quickly become intractable when scaling up to real-world problems, which remains a fundamental challenge not only in the behavioral sciences but across many other disciplines. Establishing such a bounded rational framework for sensorimotor learning and testing it in technical and biological sensorimotor applications is driving state of the art. Expected results until the end of the project include relating information-processing constraints to brain activity measured by fMRI, studying the transition between heuristic and Bayes-optimal strategies depending on resource availability, and the application of the theory to groups of human decision-makers.