SEMO: A Socio-Evolutionary Adaptive Optimization Framework for Dynamic Social Network Tie Management

TL;DR

The paper addresses how individuals navigate uncertain, dynamic social environments where forming or reinforcing ties has long-term consequences. It introduces Social-UCB, a unified framework that combines multi-armed bandits for tie formation with MDP-based long-term planning, embedded in an agent-based, evolving network model. A socio-evolutionary fitness function, along with bounded-rational update rules and UCB guarantees, enables normative policy optimization and robust exploration. Through simulations, the approach shows superior cumulative fitness and tighter network cohesion compared to baselines, offering a scalable tool for studying adaptive social behavior and informing platform designs that balance exploration and social stability.

Abstract

We propose a novel computational framework that models human social decision-making under uncertainty as an integrated Multi-Armed Bandit (MAB) and Markov Decision Process (MDP) optimization problem, in which agents adaptively balance the exploration of new social ties and the exploitation of existing relationships to maximize a socio-evolutionary fitness. The framework combines reinforcement learning, Bayesian belief updating, and agent-based simulation on a dynamic social graph, allowing each agent to use bandit-based Upper-Confidence-Bound (UCB) strategies for tie formation within an MDP of long-term social planning. We define a formal socio-evolutionary fitness function that captures both individual payoffs (e.g. shared information or support) and network-level benefits, and we derive update rules incorporating cognitive constraints and bounded rationality. Our Social-UCB algorithm, presented in full pseudocode, provably yields logarithmic regret and ensures stable exploitation via UCB-style bounds. In simulation experiments, Social-UCB consistently achieves higher cumulative social fitness and more efficient network connectivity than baseline heuristics. We include detailed descriptions of envisioned figures and tables (e.g. network evolution plots, model comparisons) to illustrate key phenomena. This integrated model bridges gaps in the literature by unifying exploration-exploitation dynamics, network evolution, and social learning, offering a rigorous new tool for studying adaptive human social behavior.

Figures (3)

• Figure 1: Conceptual Architecture Diagram. Each agent maintains a local observation of its social state, including nodes (agents) and edge weights (tie strength). A reinforcement learning module computes value estimates ($Q$-values) for exploiting known ties and UCB scores for exploring new ones. Blue arrows indicate exploratory actions based on confidence bounds; green arrows reflect exploitation choices guided by learned values. Feedback from interactions updates both the agent’s belief and the network structure. The socio-evolutionary fitness function integrates both reward and network utility to guide adaptive decision-making.
• Figure 2: Expected Learning Curves. Average cumulative fitness over time for each model. Social-UCB (blue line) demonstrates steady improvement and asymptotic convergence toward near-optimal performance. Baseline strategies—Random Walk (red) and Exploit-Only (green)—plateau early and underperform due to insufficient exploration or premature commitment. Shaded areas represent 95% confidence intervals over Monte Carlo trials.
• Figure 3: Network Evolution Over Time. Dynamics of key network properties—average degree and clustering coefficient—under different strategies. Social-UCB promotes stable and cohesive substructures (solid lines), while fragile baselines yield sparser and fragmented topologies (dashed lines). Graph snapshots at $t=0$, $t=100$, and $t=300$ (not shown here) reveal modular formation and persistent hubs in the Social-UCB condition.