Potential-Based Shaping in Hierarchical Models

• Potential-based shaping is defined as augmenting hierarchical decision models with auxiliary potential functions to systematize multi-criteria comparisons.
• It employs lexicographic ordering via aggregated potentials at different levels, balancing detailed preference capture with computational challenges.
• Algorithmic approaches include MILP formulations, recursive pruning, and dynamic programming, enabling practical applications in optimization and policy synthesis.

Potential-based shaping refers to the augmentation of multi-criteria or hierarchical decision models—particularly in settings involving preferences by levels or lexicographic importance—where auxiliary “potential” functions or structures are constructed to systematically guide comparison, inference, or optimization over alternatives. This principle is foundational in the analysis and computation of preference hierarchies, especially within hierarchical constraint logic programming (HCLP) models, multi-objective decision frameworks, preference inference in partial orders, and the synthesis of preference-driven policies in stochastic or constrained environments.

1. Foundations of Potential-Based Shaping in Hierarchical Models

The core of potential-based shaping in hierarchical models is the aggregation and ordering of evaluation criteria into a sequence of levels, each of which potentially determines the comparison between alternatives. Formally, given a set of criteria $C=\{c_1,\ldots,c_n\}$ and an unknown or partially known importance ordering, the criteria are partitioned into equivalence classes $E_1 \succ E_2 \succ \cdots \succ E_k$, reflecting their hierarchical importance. Each alternative is described by a cost-vector $\mathbf{cost}(a)$, and alternatives are compared using a lexicographic rule varying by potential at each level. The “potential” at each stage refers to the aggregate measure (e.g., sum) over criteria in that equivalence class, and is used to break ties or determine preference recursively down the hierarchy (Wilson et al., 2024).

This framework is general: it subsumes total orders (where equivalence classes are singletons), multi-level lexicographic orders, and, with more generality, arbitrary partitions reflecting known or hypothesized user priority structure.

2. Formal Structure and Key Definitions

Let $A$ denote a finite set of alternatives and $\mathcal{F} = \{f_1, \ldots, f_n\}$ a set of non-negative rational evaluation functions (e.g., costs, rewards). The hierarchical model is a sequence $H = (C_1, \ldots, C_k)$ of disjoint subsets of $\mathcal{F}$, ordered by priority. For any two alternatives $\alpha,\beta \in A$, define the induced lexicographic order: $$\alpha \preccurlyeq_H^\oplus \beta \Longleftrightarrow (\forall i, \alpha \equiv_{C_i}^\oplus \beta) \vee \exists j \text{ least, s.t. } (\forall i<j, \alpha \equiv_{C_i}^\oplus \beta) \wedge (\alpha \prec_{C_j}^\oplus \beta),$$ where the potential at level $C_i$ is aggregated via a strictly monotonic operator $E_1 \succ E_2 \succ \cdots \succ E_k$0 (typically sum), and $E_1 \succ E_2 \succ \cdots \succ E_k$1 if the potentials are equal. This structure encodes the potential-based shaping: decision making proceeds by evaluating the highest-priority group whose potential is not tied (George et al., 2024).

In practical inference and planning, this lexicographic ordering is “shaped” by the assignment of criteria to levels and by the aggregation function, constituting the scalar potential at each level.

3. Computational Complexity and Inference

Two central problems concerning potential-based shaping are preference inference (deduction) and preference consistency (PCP):

• Preference Inference: Given a set of observed pairwise preference statements, one seeks to infer whether a new preference is entailed under all hierarchical (potential-based) models compatible with observed data. With unrestricted ties (equivalence-class sizes $E_1 \succ E_2 \succ \cdots \succ E_k$2), this problem is coNP-complete. If only total orders are permitted ($E_1 \succ E_2 \succ \cdots \succ E_k$3), the problem is polynomial-time solvable (Wilson et al., 2024).
• Preference Consistency (PCP): Deciding if a set of preference statements is consistent with some hierarchical model is NP-complete for $E_1 \succ E_2 \succ \cdots \succ E_k$4, but again tractable for singleton levels or fixed partitions (George et al., 2024).

Potential-based shaping therefore directly determines the tractability of preference reasoning: as soon as the “potential structure” (level sets) allows grouping, computational hardness emerges; when the potential-shaping is strictly linear (total order), efficient algorithms are possible.

4. Algorithmic Approaches and Optimization

Potential-based shaping yields a spectrum of algorithmic strategies, leveraging the potential structure:

• MILP Formulations: Mapping the assignment of criteria to level-sets and the lex-ordered comparison constraints into integer programming problems, where variables track the assignment and potential values at each level. This approach is practical for small problems, but scalability is limited by the combinatorial nature of potential assignments (George et al., 2024).
• Recursive Pruning: Algorithms such as PC-check exploit potential-based structure by greedily extending singleton levels (maximally resolving non-strict equalities via one criterion at a time), then backtracking only when necessary. Lemmas such as strict-monotonicity ensure that once a potential at a given level breaks a tie, lower levels can be constructed independently.
• Dynamic Programming in Preference-Based Planning: In MDP-based environments, potential over temporally extended goals can be explicitly modeled as a “preference automaton,” which in turn induces potential-based shaping of product-MDP state space and the identification of Pareto-optimal policies—those whose vectors of potentials (expected satisfaction for each goal) are undominated under stochastic dominance relations (Rahmani et al., 2024).

5. Practical Implications and Expressiveness

The expressiveness of potential-based shaping in hierarchical models strictly exceeds that of simple weighted sum or fixed lexicographic approaches. Allowing sets of jointly-most-important criteria (grouped potentials) enables the representation of nuanced real-world hierarchies and incomplete priority specifications. However, this comes at a steep computational price: even a single “two-way tie” (i.e., a pair of criteria of equal priority) shifts preference inference to coNP-complete or NP-complete status (Wilson et al., 2024, George et al., 2024). This underscores the trade-off between expressive potential shaping and algorithmic efficiency.

In applied domains—such as multi-objective optimization, user preference elicitation, and sequential recommendation—potential-based shaping enables systems to align more closely with observed or elicited structures of importance, but often requires careful tuning to remain tractable.

6. Illustrative Examples and Concrete Workflows

A typical workflow under potential-based shaping involves:

1. Modeling: Define criteria and (if possible) groupings into importance levels; for unknown user hierarchies, enumerate or optimize over possible partitions.
2. Preference Encoding: Represent alternatives as vectors of potential-relevant features; translate user statements into constraints on potentials at each level.
3. Deduction/Consistency Checking: Use polynomial-time algorithms for total orders (by propagating “support vs. opposition” in the order of singled-out criteria), or recursive/MILP-based solvers for general groupings.
4. Decision-Making/Policy Synthesis: In stochastic or multi-agent settings, use product structures or automata to track fulfillment of potential-shaped objectives, assembling Pareto-frontiers in the space of potentials (Rahmani et al., 2024).
5. Interpretation: Analyze inconsistency bases to identify minimal subsets of preferences or criteria causing incompatibility, and refine potential group structure as needed.

The following table contrasts tractable and intractable cases:

Potential Grouping	Deduction/Consistency Complexity	Algorithmic Approach
Total order (singletons)	P	Sequence pruning
Fixed partition/equivalence class	P	Partition-based pruning
Arbitrary groupings (ties) $E_1 \succ E_2 \succ \cdots \succ E_k$5	coNP-complete / NP-complete	MILP / recursive search

7. Extensions and Theoretical Connections

Potential-based shaping is tightly connected with coalgebraic generalizations of choice and preference hierarchies, as in the universal choice structure framework (Galeazzi et al., 2023). Here, maximization of potential at each stage corresponds to the selection of maximal elements by choice functions, and the hierarchical composition of such functions forms the backbone of a general, non-redundant theory encompassing classical and non-classical preference relations. This suggests that potential-based shaping techniques can be abstracted and extended beyond traditional multi-criteria decision making, into settings involving higher-order choice structures, recursive learning of partial preferences, and dynamic environments with evolving hierarchies.

Potential-based shaping, therefore, occupies a central role in bridging formal preference theory, computational reasoning, and practical systems for decision support and recommendation.