Multidisciplinary Definitions and Driving Mechanisms of Preference偏好的多学科定义与驱动机制

The following content is generated by LLMs and may contain inaccuracies.

Context

Your proposed idea touches on the fundamental heterogeneity of the concept of preference (preference) across disciplines, a tension that is often obscured in interdisciplinary research. Behavioral economics aims to improve or comprehensively overhaul traditional economic theory by studying biases, inclinations, and heuristics in people’s economic decision-making, while computer science formalizes preferences as utility functions to be optimized. The free energy principle posits that systems follow a path of minimal surprise, with the brain reducing surprise or uncertainty by making predictions based on internal models and updating those models using sensory input to improve predictive accuracy. Quantum decision theory represents cognitive states and prospects as vectors in Hilbert space, the idea that systems exist in superposition of states before measurement being analogous to how our cognitive states are indeterminate before making decisions. This diversity of perspectives reflects a deeper question: are preferences revelatory (observed choices), constructive (context-dependent constructions), or metaphysical (expressions of value hierarchies)? The current dialogue among neuroscience, artificial intelligence, and normative ethics requires explicit clarification of these ontological commitments.

Key Insights

1. The Incommensurability of Driving Mechanisms Reveals the Boundaries of Modeling Assumptions
Preference-based reinforcement learning involves an agent acting according to a given policy and an expert evaluating that behavior; three distinct learning approaches include learning the policy, learning a preference model, or learning a utility function. These approaches are not interchangeable in practice: modeling human preferences as informed by regret (a measure of how far a single action deviates from the optimal decision) rather than partial rewards demonstrates that in multiple contexts, the former possesses reward function identifiability while the latter lacks this property. Heuristics are typically defined as cognitive shortcuts or rules of thumb that simplify decision-making under uncertain conditions; they represent the process of substituting a simpler problem for a difficult one, implying that “preference” may be a byproduct of metacognitive processes rather than an independent entity. A biological perspective offers another framework: under the free energy principle, biological agents act to maintain themselves within a restricted set of preferred states of the world, learning the generative model of the world and planning future actions to sustain a homeostasis that satisfies their preferences. These mechanisms—Bayesian inference, heuristic substitution, reward maximization—cannot be reduced to one another; they constitute distinct explanatory paradigms.

2. Quantum and Phenomenological Approaches Reveal the Deep Structure of Uncertainty and Contextuality
Quantum decision theory is grounded in the mathematical theory of separable Hilbert spaces, capturing superposition effects of composite prospects—multiple merged prospective actions—the theory describing entangled decision-making, the non-commutativity of successive decisions, and intentional interference. This is more than a mathematical analogy: quantum probability provides straightforward explanations for conjunction and disjunction errors and numerous other findings such as order effects in probability judgment; quantum models introduce a new fundamental concept—the compatibility and incompatibility of questions and their effects on the order of judgment. Simultaneously, in Scheler’s ethics, love is not merely an emotion but a cognitive act that recognizes values and arranges them in an ordo amoris (order of love); Scheler describes four value hierarchies—the sensory (pleasure and pain), the vital (health, vitality), the spiritual (beauty, truth, justice), and the sacred (holiness, divinity)—with the correct ordo amoris involving loving higher values over lower ones. These perspectives together suggest that preferences are not static orderings but dynamic structures that collapse at the moment of measurement/action, shaped by the value ontology of the individual or culture.

3. Interdisciplinary Integration Requires a Meta-theoretical Framework Rather Than Reductive Translation
The current gap cannot be bridged through terminological alignment but requires a framework capable of accommodating multiple causal levels. Beliefs about world states and policies are continuously updated to minimize variational free energy, wherein posterior beliefs about policies are based on expected free energy; both self-evidence and active inference entail a fundamental requirement to minimize generalized free energy or uncertainty. However, cognitive biases, heuristics, affect, and social influences all play critical roles in shaping economic choices, leading individuals' behavior to deviate from rationality; behavioral economics emphasizes how emotions interact with cognitive biases to influence decision-making. An integrative framework might resemble Scheler’s ordo amoris as “meta-preferences”—preferences about how to weigh values across different domains—or it might resemble active inference unifying perception and action under free energy minimization. The focus in military science on commander’s intent and courses of action hints at another dimension: preferences are embedded in agent-environment coupling rather than residing solely “within” the agent.

Open Questions

Q1: Does a “true” definition of preference exist, or does each discipline construct a functionally equivalent but ontologically incompatible concept? If the latter is true, should interdisciplinary research abandon the search for a unified definition in favor of constructing explicit translation protocols—analogous to dualities in physics (such as wave-particle duality)?

Q2: How does “commander’s intent” in military science relate to policy selection in computational agents? Is there an “embedded preferences” theory that views individual preferences as emergent properties within larger hierarchical systems (organizations, cultures, ecologies), thereby bridging individual and collective-level analysis?