Multi-objective CPT Optimization

• Multi-objective CPT is a decision-making framework that integrates behavioral cumulative prospect theory with multi-objective optimization to capture human risk biases.
• It applies non-linear probability weighting and S-shaped value functions to Pareto fronts in MDPs and VMP, providing tractable solutions for complex trade-offs.
• The approach combines CP-preference modeling with evolutionary algorithms to address combinatorial challenges and ensure preference-conforming, Pareto-optimal decisions.

Multi-objective Cumulative Prospect Theory (CPT) refers to the integration of cumulative prospect theory—a behavioral model for decision-making under risk and uncertainty—within the context of multi-objective optimization and decision processes. This intersection addresses both the computational and modeling challenges of incorporating empirically validated human preferences and biases into complex multi-criteria decision frameworks such as Markov decision processes (MDPs), virtual machine placement (VMP), and combinatorial optimization problems with conflicting objectives.

1. Foundations of Cumulative Prospect Theory in Multi-objective Settings

Cumulative Prospect Theory (CPT) extends expected utility theory by introducing empirically realistic components such as nonlinear probability weighting and asymmetric value functions for gains and losses. Formally, a "prospect" is represented as a set $x = ([o_1:p_1,\ldots,o_k:p_k])$, where outcomes $o_i$ occur with probabilities $p_i$. The value function $v(o)$ is typically S-shaped, and the probability-weighting function $w(p)$ reflects the empirically-observed tendency to overweight small probabilities and underweight large ones, e.g.,

$$w(p) = \frac{p^\gamma}{(p^\gamma + (1-p)^\gamma)^{1/\gamma}}, \quad \gamma\in(0,1)$$

The CPT-value aggregates decision weights $\pi_i$ and subjective utilities via

$\mathrm{CPT}(x) = \sum_{i=1}^k v(o_i)\,\pi_i(x)$

In multi-objective frameworks, these constructs must either be generalized for vector-valued returns or, more commonly, applied to Pareto-front representations or weighted reachability outcomes, as in MDPs with multiple targets (Brihaye et al., 14 May 2025).

2. Multi-objective CPT in Markov Decision Processes

In the multi-objective extension to MDPs, CPT is used to evaluate policies under weighted reachability objectives, where multiple target sets $F_1,\ldots,F_k$ each correspond to distinct outcomes or objectives. For each scheduler $\sigma$, the vector of reachability probabilities $o_i$0 constitutes a prospect, to which the CPT functional is applied: $o_i$1 where $o_i$2 is the set of achievable Pareto-optimal probability vectors. The optimal value can always be attained by a memoryless randomized strategy; deterministic policies may not suffice, as established via multi-objective reachability results (Brihaye et al., 14 May 2025).

Computationally, the problem of deciding whether the CPT-value exceeds a threshold is shown to be in EXPTIME, with fixed-parameter tractability in the number of objectives and the number of outcomes. For Markov chains (MCs), a polynomial-time algorithm exists, leveraging stationary distribution computation and direct prospect aggregation.

3. Preference Modeling via Ceteris Paribus Statements

Ceteris Paribus (CP) preference modeling formalizes decision maker (DM) priorities in multi-objective combinatorial problems by specifying strict orderings over variable assignments, "all else equal." In the Virtual Machine Placement (VMP) context, the CP structure is a set $o_i$3, where $o_i$4 indexes preference variables, e.g., VM placements. CP-dominance is defined as: $o_i$5 where $o_i$6 is the set of indices where assignments differ. A solution is "CPR-Pareto" if it is not CP-dominated by any other. Testing CP-dominance is $o_i$7, where $o_i$8 is the number of preference variables (Alashaikh et al., 2019).

This framework enables prioritization of non-objective-based, interpretable preferences, offering an alternative or complementary selection criterion to Pareto-optimality in evolutionary algorithms for multi-objective combinatorial optimization.

4. Algorithmic Approaches and Complexity

Algorithmic integration of multi-objective CPT generally takes two forms:

• Search on Pareto Frontiers: Maximum CPT-value is sought over the Pareto frontier of the multi-objective reachability problem. Approximations proceed via $o_i$9-approximation grids and non-convex Lipschitz optimization on the probability simplex (Brihaye et al., 14 May 2025).
• CP-based Evolutionary Optimization: The CP-NSGA algorithm modifies NSGA-II by applying CP-dominance as the first selector on the last nondominated front, preserving standard Pareto solutions and ensuring computational overhead remains negligible ($p_i$0 in empirical studies (Alashaikh et al., 2019)).

In combinatorial settings, such as the bi-objective cable-trench problem (Löhken et al., 2023), $p_i$1-constraint scalarizations coupled with problem-specific cutting planes are efficient for generating all non-dominated solutions, although the general enumeration problem is multi-objective intractable (number of Pareto points can grow exponentially).

Problem	Preference/Objective Model	Computational Result
MCs/MDPs (Brihaye et al., 14 May 2025)	Multi-objective CPT	PTIME (MCs), EXPTIME (MDPs)
VMP (Alashaikh et al., 2019)	CP-preference	$p_i$2 dominance, $p_i$3 overhead
Cable-Trench (Löhken et al., 2023)	Pareto/lexicographic	NP-complete, exponential Pareto-front

5. Practical Applications

• Virtual Machine Placement (VMP): CP preference modeling enables fine-grained user-specific prioritization in multi-objective placement (e.g., communication cost, energy, resource wastage), embedded in evolutionary metaheuristics to achieve solutions that are both Pareto-optimal and preference-conforming with negligible runtime penalties (Alashaikh et al., 2019).
• Markov Decision Processes (MDPs): CPT quantifies risk-aware control objectives, capturing realistic aversions and risk sensitivities in stochastic control, with computational guarantees and optimality results for memoryless randomized strategies (Brihaye et al., 14 May 2025).
• Combinatorial Network Design: Bi-objective formulations, such as in cable-trench problem instances, highlight the necessity of enumerating non-supported solutions distinct from weighted-sum optima, reflecting complex trade-offs in multi-criteria infrastructure planning (Löhken et al., 2023).

6. Empirical Observations and Extensions

In benchmark studies for VMP under CP-NSGA, preference satisfaction metrics ("weighted-flips") were reduced significantly compared to standard implementations, with higher preference uptake as the number of CP-specified variables increases. The approach preserves Pareto-front diversity and solution quality.

In the cable-trench context, the interplay of graph structure and edge-weight correlation drives the cardinality and tractability of the non-dominated set. Hybrid scalarization approaches can avoid weakly-efficient solutions, and combinatorial explosion necessitates approximation algorithms for larger problem instances.

Multi-objective CPT naturally extends to broader variants, including mean-payoff objectives, capacity constraints, and multi-commodity extensions, through adapted scalarization, branch-and-cut, or heuristic schemes.

7. Connections to Related Multi-objective Frameworks

Multi-objective CPT situates itself at the intersection of behavioral decision theory, combinatorial optimization, and algorithmic game theory. While standard multi-objective optimization employs Pareto efficiency as the main selection criterion, multi-objective CPT and CP-preference frameworks allow for direct modeling of non-linear, possibly non-convex human stakeholder priorities and context-dependent trade-offs. This alignment with empirically observed decision behaviors facilitates more realistic automated decision-support in both stochastic and deterministic multi-objective regimes.

References for formal models, algorithms, and complexity proofs are found in (Brihaye et al., 14 May 2025, Alashaikh et al., 2019), and (Löhken et al., 2023).