Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents

Wouter M. Kouw

doi:10.1007/978-3-031-77138-5_13

Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents

Wouter M. Kouw (Corresponding author)

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

Abstract

In nature, active inference agents must learn how observations of the world represent the state of the agent. In engineering, the physics behind sensors is often known reasonably accurately and measurement functions can be incorporated into generative models. When a measurement function is non-linear, the transformed variable is typically approximated with a Gaussian distribution to ensure tractable inference. We show that Gaussian approximations that are sensitive to the curvature of the measurement function, such as a second-order Taylor approximation, produce a state-dependent ambiguity term. This induces a preference over states, based on how accurately the state can be inferred from the observation. We demonstrate this preference with a robot navigation experiment where agents plan trajectories.

Original language	English
Title of host publication	Active Inference
Subtitle of host publication	5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers
Editors	Christopher L. Buckley, Daniela Cialfi, Pablo Lanillos, Riddhi J. Pitliya, Noor Sajid, Hideaki Shimazaki, Tim Verbelen, Martijn Wisse
Place of Publication	Cham
Publisher	Springer
Pages	195-208
Number of pages	13
ISBN (Electronic)	978-3-031-77138-5
ISBN (Print)	978-3-031-77137-8
DOIs	https://doi.org/10.1007/978-3-031-77138-5_13
Publication status	Published - 31 Dec 2024
Event	5th International Workshop on Active Inference, IWAI 2024 - Oxford, UK, Oxford, United Kingdom Duration: 9 Sept 2024 → 11 Sept 2024

Publication series

Name	Communications in Computer and Information Science (CCIS)
Volume	2193
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	5th International Workshop on Active Inference, IWAI 2024
Abbreviated title	IWAI 2024
Country/Territory	United Kingdom
City	Oxford
Period	9/09/24 → 11/09/24

Funding

The author gratefully acknowledges financial support from the Eindhoven Artificial Intelligence Systems Institute (EAISI) at TU Eindhoven.

Keywords

Active inference
Free energy minimization
Bayesian filtering
Non-linear sensing
Control system
Planning
Navigation
Control systems

Access to Document

10.1007/978-3-031-77138-5_13

https://arxiv.org/pdf/2409.01974

Cite this

Kouw, W. M. (2024). Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents. In C. L. Buckley, D. Cialfi, P. Lanillos, R. J. Pitliya, N. Sajid, H. Shimazaki, T. Verbelen, & M. Wisse (Eds.), Active Inference: 5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers (pp. 195-208). (Communications in Computer and Information Science (CCIS); Vol. 2193). Springer. https://doi.org/10.1007/978-3-031-77138-5_13

Kouw, Wouter M. / Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents. Active Inference: 5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers. editor / Christopher L. Buckley ; Daniela Cialfi ; Pablo Lanillos ; Riddhi J. Pitliya ; Noor Sajid ; Hideaki Shimazaki ; Tim Verbelen ; Martijn Wisse. Cham : Springer, 2024. pp. 195-208 (Communications in Computer and Information Science (CCIS)).

@inproceedings{9ba4c96a27e34d7d8e3494e23c200498,

title = "Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents",

abstract = "In nature, active inference agents must learn how observations of the world represent the state of the agent. In engineering, the physics behind sensors is often known reasonably accurately and measurement functions can be incorporated into generative models. When a measurement function is non-linear, the transformed variable is typically approximated with a Gaussian distribution to ensure tractable inference. We show that Gaussian approximations that are sensitive to the curvature of the measurement function, such as a second-order Taylor approximation, produce a state-dependent ambiguity term. This induces a preference over states, based on how accurately the state can be inferred from the observation. We demonstrate this preference with a robot navigation experiment where agents plan trajectories.",

keywords = "Active inference, Free energy minimization, Bayesian filtering, Non-linear sensing, Control system, Planning, Navigation, Control systems",

author = "Kouw, \{Wouter M.\}",

year = "2024",

month = dec,

day = "31",

doi = "10.1007/978-3-031-77138-5\_13",

language = "English",

isbn = "978-3-031-77137-8",

series = "Communications in Computer and Information Science (CCIS)",

publisher = "Springer",

pages = "195--208",

editor = "Buckley, \{Christopher L.\} and Daniela Cialfi and Pablo Lanillos and Pitliya, \{Riddhi J.\} and Noor Sajid and Hideaki Shimazaki and Tim Verbelen and Martijn Wisse",

booktitle = "Active Inference",

address = "Germany",

note = "5th International Workshop on Active Inference, IWAI 2024, IWAI 2024 ; Conference date: 09-09-2024 Through 11-09-2024",

}

Kouw, WM 2024, Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents. in CL Buckley, D Cialfi, P Lanillos, RJ Pitliya, N Sajid, H Shimazaki, T Verbelen & M Wisse (eds), Active Inference: 5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers. Communications in Computer and Information Science (CCIS), vol. 2193, Springer, Cham, pp. 195-208, 5th International Workshop on Active Inference, IWAI 2024, Oxford, United Kingdom, 9/09/24. https://doi.org/10.1007/978-3-031-77138-5_13

Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents. / Kouw, Wouter M. (Corresponding author).
Active Inference: 5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers. ed. / Christopher L. Buckley; Daniela Cialfi; Pablo Lanillos; Riddhi J. Pitliya; Noor Sajid; Hideaki Shimazaki; Tim Verbelen; Martijn Wisse. Cham: Springer, 2024. p. 195-208 (Communications in Computer and Information Science (CCIS); Vol. 2193).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents

AU - Kouw, Wouter M.

PY - 2024/12/31

Y1 - 2024/12/31

N2 - In nature, active inference agents must learn how observations of the world represent the state of the agent. In engineering, the physics behind sensors is often known reasonably accurately and measurement functions can be incorporated into generative models. When a measurement function is non-linear, the transformed variable is typically approximated with a Gaussian distribution to ensure tractable inference. We show that Gaussian approximations that are sensitive to the curvature of the measurement function, such as a second-order Taylor approximation, produce a state-dependent ambiguity term. This induces a preference over states, based on how accurately the state can be inferred from the observation. We demonstrate this preference with a robot navigation experiment where agents plan trajectories.

AB - In nature, active inference agents must learn how observations of the world represent the state of the agent. In engineering, the physics behind sensors is often known reasonably accurately and measurement functions can be incorporated into generative models. When a measurement function is non-linear, the transformed variable is typically approximated with a Gaussian distribution to ensure tractable inference. We show that Gaussian approximations that are sensitive to the curvature of the measurement function, such as a second-order Taylor approximation, produce a state-dependent ambiguity term. This induces a preference over states, based on how accurately the state can be inferred from the observation. We demonstrate this preference with a robot navigation experiment where agents plan trajectories.

KW - Active inference

KW - Free energy minimization

KW - Bayesian filtering

KW - Non-linear sensing

KW - Control system

KW - Planning

KW - Navigation

KW - Control systems

UR - http://www.scopus.com/inward/record.url?scp=85215801079&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-77138-5_13

DO - 10.1007/978-3-031-77138-5_13

M3 - Conference contribution

SN - 978-3-031-77137-8

T3 - Communications in Computer and Information Science (CCIS)

SP - 195

EP - 208

BT - Active Inference

A2 - Buckley, Christopher L.

A2 - Cialfi, Daniela

A2 - Lanillos, Pablo

A2 - Pitliya, Riddhi J.

A2 - Sajid, Noor

A2 - Shimazaki, Hideaki

A2 - Verbelen, Tim

A2 - Wisse, Martijn

PB - Springer

CY - Cham

T2 - 5th International Workshop on Active Inference, IWAI 2024

Y2 - 9 September 2024 through 11 September 2024

ER -

Kouw WM. Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents. In Buckley CL, Cialfi D, Lanillos P, Pitliya RJ, Sajid N, Shimazaki H, Verbelen T, Wisse M, editors, Active Inference: 5th International Workshop, IWAI 2024, Oxford, UK, September 9–11, 2024, Revised Selected Papers. Cham: Springer. 2024. p. 195-208. (Communications in Computer and Information Science (CCIS)). doi: 10.1007/978-3-031-77138-5_13

Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents

Abstract

Publication series

Conference

Funding

Keywords

Access to Document

Other files and links

Fingerprint

CRT-STA-FEP-walker LWAI: Learning to Walk by Active Inference

Cite this

Planning to avoid ambiguous states through Gaussian approximations to non-linear sensors in active inference agents

Abstract

Publication series

Conference

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Projects

CRT-STA-FEP-walker LWAI: Learning to Walk by Active Inference

Cite this