Traditional game theory operates under a foundational assumption: agents are strictly self-interested. In cooperative settings like coalition formation or public goods games, an individual's utility is typically modeled as a direct function of their own immediate gains or the specific group they join. However, real-world social structures do not operate in a vacuum of pure selfishness. People care about the well-being of their friends, colleagues, and peers.
To bridge this gap, a growing body of research focuses on altruistic preference modeling. This paradigm redefines utility functions by integrating an agent's personal payoffs with the welfare of their social network. By formalizing how social context alters strategic decisions, this framework explains why human coalitions form, stabilize, or dissolve in ways that purely selfish models fail to predict.
The Core Mechanism: Socially-Aware Utility
The fundamental shift in altruistic preference modeling lies in how an agent's utility is constructed. In a standard hedonic game—where players split into disjoint coalitions—a player $i$'s utility depends solely on the members of their own coalition. As noted by Iyad Rahwan et al. in Pareto Optimality in Abstract Argumentation and Simina Brânzei et al. in Social Distance Games, ignoring interpersonal relationships fails to capture scenarios where individuals derive happiness from intrinsic group properties like proximity, affinity, or the satisfaction of others.
To formalize this "social context," researchers modify the utility function to be a weighted combination of personal valuation and peer welfare. For instance, Gianpiero Monaco et al. introduced Social Context Hedonic Games (SCHGs) in Hedonic Games with Social Context, where a player's payoff is explicitly influenced by the happiness of their friends.
Mathematically, this can be expressed through an altruism coefficient or parameter. In Additively Separable Hedonic Games with Social Context, Gianpiero Monaco et al. parameterize this relationship so that an agent's utility balances their direct valuation of their own coalition against the utilities their friends receive from their respective coalitions. Similarly, in networked public goods scenarios, Arnab Maiti et al. in On Binary Networked Public Goods Game with Altruism overlay an "altruism network" on top of the standard game graph, forcing players to internalize the costs and benefits experienced by their neighbors.
How Agents Aggregate Peer Welfare
Not all altruism is the same. How do agents calculate the "collective happiness" of their friends? The literature distinguishes between two primary aggregation strategies:
1. Average-Based (Utilitarian) Altruism
In this model, an agent considers the average or total utility of their social circle. For example, Po-An Chen et al. in The Robust Price of Anarchy of Altruistic Games model a player's perceived cost as a weighted average of their personal cost and the total social cost. This utilitarian approach assumes that an agent is satisfied as long as their friends are doing well on average, even if a few friends are highly dissatisfied.
2. Minimum-Based (Egalitarian) Altruism
Conversely, an agent might be governed by the "weakest link" in their social network. Introduced by Alessandra Wiechers et al. in Stability in Minimization-Based Altruistic Hedonic Games, this approach dictates that an individual's altruistic utility is bounded by the least satisfied friend. This models a social reality where a group's collective cohesion is limited by its most vulnerable or unhappy member.
To capture the spectrum between these extremes, Anna Maria Kerkmann et al. in Altruistic Hedonic Games classify altruism into distinct tiers: selfish-first (where altruism only breaks ties), equal weight, and friend-centric decision-making (where peer satisfaction overrides personal preference).
The Local vs. Global Altruism Paradox
A critical breakthrough in this field involves the scope of an agent's concern. Early models relied on local altruism, meaning an agent only cared about the well-being of friends who resided within their immediate coalition.
However, local altruism introduces a severe logical inconsistency. If an agent decides to leave Coalition A to join Coalition B, a local altruism model dictates that they instantly stop caring about the friends they left behind in Coalition A. This mathematical artifact can cause agents to make strategic moves that objectively harm both themselves and their entire social network, violating the fundamental economic principle of unanimity.
To resolve this, Anna Maria Kerkmann et al. in Altruism in Coalition Formation Games (ijcai 2020) and Altruism in Coalition Formation Games (arxiv 2023) proposed global altruism. Under global altruism, an agent's utility accounts for the welfare of their friends across the entire system, regardless of group boundaries.
Local Altruism: [Agent i + Friend j] in Coalition A --> i cares about j
Agent i moves to Coalition B --> i ceases to care about j
Global Altruism: [Agent i] in Coalition B -----------_> i still cares about j
[Friend j] in Coalition A _________/ across coalition boundaries
By decoupling social ties from organizational boundaries, global altruism ensures that preference changes satisfy the unanimity property: an agent will only move if it does not cause a net systemic collapse in their social circle's welfare.
System Dynamics: Stability, Efficiency, and Complexity
Does adding altruism make systems more stable and efficient, or does it complicate them? The answer is a nuanced mix of both.
Stability and Potential Functions
Altruism frequently acts as a stabilizing force. In many social context games, the alignment of individual goals with peer welfare allows researchers to construct an exact potential function. As demonstrated by the authors in Hedonic Games with Social Context, the existence of a potential function guarantees that decentralized groups will naturally converge to a stable Nash equilibrium rather than cycling endlessly.
Efficiency: Windfalls and Prices
The impact of altruism on system efficiency is dual-faceted:
- The Windfall of Friendship (WoF): Dominic Meier et al. in On the Windfall and price of friendship: Inoculation strategies on social networks show that caring about neighbors generally mitigates the inefficiencies of purely selfish behavior, driving down the Price of Anarchy.
- The Price of Friendship: Conversely, the same authors note that social ties can significantly slow down convergence times to equilibrium, as agents must weigh complex interpersonal ripples before making a move.
Computational Complexity
While altruism can guarantee that stable states exist, actually finding or verifying them becomes computationally daunting. Because an individual's utility depends on an interconnected web of peer payoffs, calculating stable configurations introduces high-dimensional dependencies. Jörg Rothe et al. in Solving Four Open Problems about Core Stability in Altruistic Hedonic Games and Anna Maria Kerkmann et al. in Popularity and Strict Popularity in Altruistic Hedonic Games and Minimum-Based Altruistic Hedonic Games proved that verifying core stability and popularity in these settings is coNP-complete, regardless of whether average-based or minimum-based functions are used.
Real-World Applications
Altruistic preference modeling extends far beyond abstract game theory into practical distributed systems:
- Federated Learning: In collaborative machine learning, optimizing for "egalitarian fairness" can sometimes degrade performance for data-rich clients, tempting them to leave the system. Jiashi Gao et al. in Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors demonstrate that modeling external altruistic behaviors explains why these "whale" clients choose to remain in the federation, stabilizing the learning ecosystem.
- Multi-Agent Negotiation: In complex automated negotiations, Tamara C. P. Florijn et al. in A Survey on One-To-Many Negotiation: A Taxonomy of Interdependency highlight that future negotiation frameworks must incorporate the utility functions of all participating agents to accurately map out collaborative interdependencies.
By mathematically formalizing empathy, altruistic preference modeling provides the analytical tools necessary to design stable, efficient, and socially-aware multi-agent systems.
Key Foundational Work
The Robust Price of Anarchy of Altruistic Games (2011) — contributed 3 concept elements to this direction.
Pareto Optimality in Abstract Argumentation (2008) — contributed 1 concept elements to this direction.
Social Distance Games (2011) — contributed 1 concept elements to this direction.