Strategically Robust Aggregative Games

Strategically Robust Aggregative Games: A Detailed Analysis

Introduction and Motivation

This paper introduces a novel, robust equilibrium concept for multi-agent convex aggregative games, motivated by decision-making problems with endogenous uncertainty arising from incomplete information, limited computation, or bounded rationality on the part of interacting agents. The core contribution is the definition and analysis of strategically robust Wardrop equilibria (SRWE), in which each agent explicitly guards against neighborhood deviations from the emergent aggregate population behavior, operationalized through optimal transport-based ambiguity sets constructed via the Wasserstein metric.

Traditionally, robust and distributionally robust game theory has assumed exogenous uncertainty, modeling disturbances or adversarial perturbations external to the game's structure. However, in many practical settings—traffic networks, energy and charging markets, etc.—the greatest uncertainty arises from the strategic behavior of other agents, which is a function of their own actions. This paper advances the state of the art by shifting the focus from modeling individual sources of uncertainty to robustifying agents' objectives directly against fluctuations in the aggregate population behavior.

Equilibrium Concept: Strategically Robust Wardrop Equilibrium

Given an $N$-player convex aggregative game, each agent $i$ chooses $x^i$ from a convex set $\mathcal{X}^i$ to minimize a cost $J^i(x^i, \sigma)$, where $\sigma$ is the (typically average) aggregate of all players' decisions. The SRWE is defined as a fixed point where every agent solves a distributionally robust optimization (DRO) problem: minimizing the worst-case expected cost over an ambiguity set of aggregate behaviors lying within a Wasserstein ball of radius $\varepsilon$ around the nominal aggregate $\sigma(x)$.

Formally, for each agent $i$,

$$x_W^i \in \argmin_{x^i \in \mathcal{X}^i} \sup_{\mu \in \mathcal{B}_\varepsilon(\delta_{\sigma(x_W)})} \mathbb{E}_{\sigma \sim \mu} [J^i(x^i, \sigma)],$$

where $i$0 denotes the set of probability distributions within Wasserstein distance $i$1 of the Dirac distribution on $i$2.

Key properties:

• Interpolation: By varying $i$3, the equilibrium interpolates between classical Wardrop equilibrium ($i$4) and pure security strategies ($i$5).
• Guaranteed Existence: Existence of pure SRWEs is proven under standard convexity and continuity assumptions, using Kakutani's fixed-point theorem, leveraging set-valued analysis tools due to the non-smoothness induced by the DRO objective.
• Endogenous Robustification: The ambiguity set is centered on the emergent aggregate, not formed a priori nor as a function of arbitrary external noise, capturing intrinsic strategic uncertainty.

Tractable Reformulation and Computational Aspects

The main computational challenge is that the agent's optimization in SRWE is infinite-dimensional due to the supremum over probability distributions (the DRO term). To address this, the authors show that the SRWE problem can be reformulated as an equilibrium in an augmented, finite-dimensional convex aggregative game:

• Reformulation via Duality: By applying strong duality for Wasserstein-based DRO, the infinite-dimensional maximization reduces to an explicit maximization over the aggregate space, and for suitable cost concavity, further to a convex program via convex conjugate representations.
• Augmented Action Spaces: Each player's decision variable is extended to include a scalar dual variable capturing worst-case transport cost.
• Algorithmic Implications: This reduction enables the use of standard equilibrium-seeking methods—such as proximal best-response algorithms—for computing SRWEs efficiently with existing convex optimization solvers.

Application: Electric Vehicle Charging and Coordination-Via-Robustification

The framework is instantiated in the context of EV charging, where a population of agents must schedule charging profiles subject to local constraints and an aggregate cost determined by a price function dependent on total load. This setting is a standard benchmark in aggregative games literature due to its practical importance and the presence of strategic coupling.

Key empirical findings:

• Effect of Robustness Level: As $i$6 increases, equilibrium charging profiles flatten, reducing peak demand and exposure to high aggregate costs. For moderate $i$7, coordination emerges that smooths aggregate demand.
• Robustness against Perturbations: When the realized aggregate demand deviates from equilibrium (e.g., due to unmodeled entry/external shocks), SRWE policies yield lower worst-case and average costs compared to standard Wardrop equilibria. This confirms that SRWEs systematically mitigate adverse effects of endogenously generated uncertainty.
• Coordination-Via-Robustification Effect: Notably, for carefully chosen robustification levels, the equilibrium aligns with the system's social optimum, driving the price of anarchy to unity. Thus, robustness demand can indirectly induce global coordination without explicit incentive alignment mechanisms.

Theoretical and Practical Implications

The theoretical contributions, bridging optimal-transport DRO with convex game equilibrium, provide:

• An equilibrium concept that manages both stochastic and deterministic aggregate deviations via endogenous robustification;
• Existence results under minimal convexity and continuity requirements;
• Tractable, algorithmically practical reformulations suitable for large-scale systems.

Practically, the "coordination-via-robustification" effect demonstrates that introducing endogenous robustness can improve welfare beyond the classic tradeoffs of robust optimization. Indeed, system designers might employ calibrated robustness constraints either as direct design interventions or as policy suggestions to induce socially optimal equilibria, especially in large-scale or infrastructure systems (EV networks, smart grids, congested transportation, etc.) where collective uncertainty is unavoidable and external enforcement of cooperation is impractical.

Future Directions

Future research avenues include:

• Decentralized Algorithm Design: While proximal-response approaches are natural, systematic convergence guarantees and scalable, decentralized protocols for SRWE computation remain open.
• Characterization of Coordination Effects: Theoretical understanding of when and why robustification leads to social optimality—across general game classes—is undeveloped.
• Extension to Nonconvex and Heterogeneous Games: Analysis and computation for settings with nonconvexities, nonsmoothness, or heterogeneity in players’ robustness parameters could yield richer behavioral phenomena and practical insights.

Conclusion

This work synthesizes optimal transport-based distributional robustness and convex aggregative game theory to propose a robust equilibrium concept with strong existence and computational guarantees. Empirical evaluation in EV charging demonstrates practical utility and the emergence of a coordination-via-robustification phenomenon, raising significant implications for robust multi-agent system design and the broader understanding of robustness-induced efficiency in strategic environments.