Abstract
This research explores the counterintuitive phenomenon where a diverse team of individually "weaker" agents can outperform a uniform team of "stronger" agents. By modeling agent performance across various "world states," we demonstrate that diversity is a critical component of team strength. We propose and evaluate several voting rules for agent coordination, showing that strategic tie-breaking significantly enhances the performance of diverse ensembles. Our findings are validated through synthetic experiments and high-complexity games of Computer Go.
1. Introduction
In multi-agent systems, forming an effective team often involves selecting the "best" available agents. However, traditional models like Condorcet’s Jury Theorem assume agents have a static probability of being correct, failing to account for situational expertise. In complex domains like Go, an agent might be exceptionally strong in certain board configurations but weak in others. This paper investigates whether a diverse team of niche experts can surpass a uniform team of general experts.
2. Methodology
We model the environment as a set of world states ($\Omega$). Each agent $i$ is characterized by a probability distribution function (pdf) over possible actions for every state $s \in \Omega$.
2.1 Metrics and Definitions
- Team Strength: Defined as the weighted average of expected utility across all world states.
- Diversity Metric: Quantified using the Hellinger Distance between the action pdfs of different agents.
- Voting Rules: We evaluate three plurality-based rules for breaking ties:
- Simple: Random tie-breaking.
- Static: Ties are broken in favor of the agent with the highest overall strength.
- Optimal: Ties are broken in favor of the agent that is strongest in the specific current world state.
3. Experimental Setup
The model was evaluated using two distinct datasets:
- Synthetic Data: 1,200 random teams of 4 agents were generated using the Quantal Response (QR) model to analyze the relationship between diversity and performance.
- Computer Go: We utilized engines including Fuego 1.1, GnuGo 3.8, Pachi 9.01, and MoGo 3, conducting 1,000 games of 9x9 Go per configuration.
4. Results and Discussion
4.1 The Diversity Multiplier
Regression analysis on synthetic data indicates that diversity and individual strength both contribute positively to team performance. For a fixed level of individual strength, increasing team diversity leads to higher overall success. This "Niche Expert" effect allows weaker agents to compensate for the blind spots of stronger ones.
4.2 Performance in Computer Go
A diverse team of six agents achieved an 18.7% higher win rate against a single Fuego engine compared to a baseline. Notably, this diverse team performed 11% better than a uniform team composed solely of Fuego copies. This suggests that uniformity breeds shared errors; if one agent fails in a specific state, its identical copies are likely to repeat the mistake.
4.3 Impact of Voting Rules
The "Static" voting rule—favoring the overall strongest agent during ties—improved performance by 12.7% over simple plurality. While the "Optimal" rule is theoretically superior, the "Static" rule provides a practical and effective approximation when state-specific distributions are unknown.
5. Conclusion and Future Work
Diversity is a vital asset in multi-agent coordination. While this study proves its value, future research should focus on automated state analysis to identify niches where specific agents excel. Developing heuristics to predict state-dependent expertise could further bridge the gap between "Static" and "Optimal" voting strategies.
Summary updated