A Minimal Functional Causal Self-Model in an Artificial Survival Agent Embodied Outcome Prediction, Personal Learning, and Distributed Action Selection

Abstract A functional self-model can be investigated through what it does rather than through linguistic self-description: an agent represents its own condition, estimates how candidate actions will change that condition, revises estimates from experienced consequences, and uses those estimates in subsequent behavior. We implement this operational definition in SISE-N, a digital survival agent with bodily energy and health, partial environmental sensing, autobiographical food and route memory, irreversible mortality, and distributed competition between foraging, avoidance, exploration, and recovery drives. A life-local action-conditioned predictor learns energy change, realized recovery, and delayed injury solely from executed actions and matured feedback. Its predictions modulate candidate motor inputs without a central self-arbiter. We examine two matched-map intervention studies with a common frozen population action-value model: G1.334 compares predictive learning without behavioral access (shadow), intact predictive readout (active), and action–prediction misassignment (shuffled); G1.335 compares shadow, legacy injury-magnitude prediction, and an event-probability × severity predictor. Across nine map seeds, 27 lives generated 116,313 matched environment–decision transitions. Independent first-divergence audits identified behavioral effects of predictive readout in all eight G1.334 intervention pairs and all ten G1.335 pairs. In G1.335, personal energy forecasts achieved mean absolute errors of 0.329–0.337 versus 0.474–0.479 for a fixed weak prior. Delayed-injury event probabilities did not improve on a cue-based prior: for the calibrated agent, Brier scores were 0.05221 and 0.05189, respectively. Observed survival differences varied across maps and interventions. The experiments establish an executable, intervenable self-prediction loop and distinguish behavioral causal efficacy from predictive calibration and ecological benefit. The resulting framework makes minimal functional self-modeling testable within artificial life. Keywords: artificial life; functional self-model; embodied prediction; homeostasis; distributed action selection; causal intervention; autobiographical memory

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24295433
Primary Topic
Embodied and Extended Cognition
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Minimal Functional Causal Self-Model in an Artificial Survival Agent Embodied Outcome Prediction, Personal Learning, and Distributed Action Selection

Jizhang Huang, Huang
1 links
Zenodo (CERN European Organization for Nuclear Research)
Embodied and Extended Cognition
preprint

A Minimal Functional Causal Self-Model in an Artificial Survival Agent Embodied Outcome Prediction, Personal Learning, and Distributed Action Selection

Jizhang Huang, Huang
preprint en

Abstract

Abstract A functional self-model can be investigated through what it does rather than through linguistic self-description: an agent represents its own condition, estimates how candidate actions will change that condition, revises estimates from experienced consequences, and uses those estimates in subsequent behavior. We implement this operational definition in SISE-N, a digital survival agent with bodily energy and health, partial environmental sensing, autobiographical food and route memory, irreversible mortality, and distributed competition between foraging, avoidance, exploration, and recovery drives. A life-local action-conditioned predictor learns energy change, realized recovery, and delayed injury solely from executed actions and matured feedback. Its predictions modulate candidate motor inputs without a central self-arbiter. We examine two matched-map intervention studies with a common frozen population action-value model: G1.334 compares predictive learning without behavioral access (shadow), intact predictive readout (active), and action–prediction misassignment (shuffled); G1.335 compares shadow, legacy injury-magnitude prediction, and an event-probability × severity predictor. Across nine map seeds, 27 lives generated 116,313 matched environment–decision transitions. Independent first-divergence audits identified behavioral effects of predictive readout in all eight G1.334 intervention pairs and all ten G1.335 pairs. In G1.335, personal energy forecasts achieved mean absolute errors of 0.329–0.337 versus 0.474–0.479 for a fixed weak prior. Delayed-injury event probabilities did not improve on a cue-based prior: for the calibrated agent, Brier scores were 0.05221 and 0.05189, respectively. Observed survival differences varied across maps and interventions. The experiments establish an executable, intervenable self-prediction loop and distinguish behavioral causal efficacy from predictive calibration and ecological benefit. The resulting framework makes minimal functional self-modeling testable within artificial life. Keywords: artificial life; functional self-model; embodied prediction; homeostasis; distributed action selection; causal intervention; autobiographical memory

Zenodo (CERN European Organization for Nuclear Research)
Embodied and Extended Cognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon