Journal of Pioneering Artificial Intelligence Research

Abstract

Autonomous decision-making systems commonly assume that the objective to be pursued is predefined, focusing computational effort on generating action sequences or policies that satisfy a given goal. This work introduces a deterministic framework in which goal selection itself is treated as a primary computational process rather than an external input. The proposed approach operates over a structured internal goal library, referred to as the identity layer, which defines a bounded set of admissible objectives together with persistent constraints that remain active across all operating conditions. At each time step, the system evaluates its current context, composed of environmental conditions, internal state variables, and operational role, and maps this context to a set of constraints that restrict the admissibility of goals. Instead of evaluating or ranking all candidate goals, the framework eliminates those that violate either identity-level constraints or context-derived constraints, producing a reduced set of admissible goals. If a single goal remains, it is selected directly; if multiple admissible goals persist, a deterministic residual resolution mechanism is applied using simple ordering or priority rules within the reduced set. In cases where no admissible goal remains, predefined safe-state objectives are activated. This formulation shifts decision-making from optimization and stochastic exploration to constraint-driven reduction, ensuring that all selected goals are consistent with both internal structure and external conditions. The approach does not rely on reward functions, probabilistic policies, or global search, and therefore maintains low computational complexity and high interpretability. It is designed to operate as a pre-planning layer for existing systems such as Goal-Oriented Action Planning, providing a deterministic mechanism for generating context-consistent goals prior to action planning. Compared with reinforcement learning and optimization-based methods, the proposed framework emphasizes admissibility over optimality, making it particularly suitable for safety-critical, real-time, and resource-constrained applications. By structuring goal selection as a process of elimination under constraints, the framework provides a unified and extensible basis for autonomous decision-making in environments where consistency, predictability, and bounded behavior are essential. Agents operate with evolving behavioural libraries that are updated through experience while remaining constrained by objective and governing principles. This enables swarm-like coordination under 2

minimal supervision, where decision-making emerges through constraint-consistent elimination over a dynamically maintained action space. A self-check mechanism enforces internal consistency by evaluating admissible actions against recent behaviour and operational intent, ensuring coherent and stable autonomous operation.