Distributed Endogenous Satisficing: Autonomous Action Selection under Irreversible Mortality and Survival Constraints
Year: 2026
Type: preprint
Abstract: A survival agent must regulate several competing demands while interacting with an environment that can injure or exhaust it. Selecting an action by a single global acceptance criterion is one possible organization of this process; another is to let locally evaluated needs compete for control of action. We examine these alternatives in SISE-N, a two-dimensional digital survival system with simulated energy, health, food acquisition, local sensing, individually acquired memories, and irreversible termination at death. Four controllers form a two-by-two comparison: bounded central candidate acceptance with a fixed or body-dependent aspiration threshold, and distributed motor competition with fixed or body-dependent local channel gains. All four inherit the same frozen population Q model and are evaluated on 12 paired, previously untested seven-food maps, with a 5,000-step observation limit. The evaluation contains 48 lives and 143,847 executed environment steps. Mean observed survival is 2,476.7 steps for central fixed acceptance, 2,011.6 for central dynamic acceptance, 3,731.9 for distributed fixed competition, and 3,767.1 for distributed dynamic competition. The distributed dynamic controller exceeds the central dynamic controller by 1,755.5 paired steps (exploratory map-bootstrap 95% interval: 937.7–2,616.9; exact paired sign-flip p = .0029), whereas introducing dynamic gains within the distributed architecture changes mean survival by only 35.2 steps. Logged first-action divergences and an independent stepwise audit establish that the decision mechanisms are operational, not merely descriptive labels. A preceding resource-stratified experiment supplies complementary ecological context but is analyzed separately because its training branches diverged. The findings identify distributed need competition as a viable organization of survival-oriented action in this simulation and show why architectural effects must be separated from aspiration-threshold effects. Keywords: artificial life; distributed action selection; homeostasis; satisficing; embodied agents; irreversible mortality
Authors: Huang, Jiye, Huang, Jizhang
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234764952
