Beyond Task Performance: A Metric-Based Analysis of Sequential Cooperation in Heterogeneous Multi-Agent Destructive Foraging

Technical Summary: Beyond Task Performance: A Metric-Based Analysis of Sequential Cooperation in Heterogeneous Multi-Agent Destructive Foraging

1. Problem Statement

The paper addresses a critical gap in the study of Multi-Agent Foraging (MAF): the lack of general-purpose metrics to quantify cooperation in heterogeneous systems. While most existing research focuses on "task performance" (e.g., how much food was collected or how fast the task was completed), these metrics fail to explain how agents coordinate, how much they depend on one another, or how robust their cooperation is to failures.

The authors frame this within a sequential, heterogeneous, destructive foraging scenario. In this setting:

• Heterogeneity: Agents have different roles (Scouts vs. Foragers) with asymmetric sensing and actuation capabilities.
• Sequential Dependency: One team (Scouts) must discover resources to inform the second team (Foragers), who then remove them.
• Partial Observability & Dynamics: Agents have limited vision, and resources move over time due to environmental factors (wind/currents), requiring constant model updates.

2. Methodology

The authors propose a multi-level evaluation framework consisting of a novel suite of metrics categorized into three tiers:

• Primary Metrics (System-level Performance): Standard indicators like Percentage of Target Achieved (PTA), Root Mean Squared Error (RMSE) for environmental modeling, Normalized Time to X (NTX) for temporal efficiency, Throughput, and Idleness Metrics (Mean Idleness and Idleness Reduction Rate) to measure spatial freshness.
• Inter-team Metrics (Coordination & Dependency): These measure the "glue" between teams:
  • Discovery-to-Service Latency (DSL): The time gap between discovery by a scout and removal by a forager.
  • Inter-Team Temporal Lag (ITL): The synchronization of peak performance between teams.
  • Cooperative Success Ratio (CSR): The proportion of tasks completed specifically through inter-team cooperation.
  • Cooperation Sensitivity under Stochastic Corruption (CSSC): A novel metric that measures how much the performance of Team A degrades when Team B is subjected to increasing levels of random (corrupted) behavior.
• Intra-team Metrics (Internal Dynamics): These assess how agents within a single team interact:
  • Gini Index of Contributions: Measures workload fairness/inequality.
  • Coverage Overlap (CO): Quantifies spatial redundancy.
  • Marginal Contribution (MC): Measures the impact of an individual agent by observing performance degradation when that agent is removed (ablation).

Experimental Setup: The metrics were validated in a simulated aquatic cleaning scenario (plastic waste removal) using three algorithms: Double Deep Q-Learning (DDQL), a Greedy heuristic, and a Lévy Walk stochastic approach.

3. Key Contributions

1. Formalization of a Sequential Heterogeneous Foraging Problem: A mathematical framework that explicitly captures temporal dependencies and role specialization under partial observability.
2. A Systematic Suite of Cooperation Metrics: A transferable set of metrics designed to move the field beyond "black-box" performance evaluation toward a diagnostic understanding of cooperative mechanisms.
3. Diagnostic Experimental Analysis: A demonstration of how these metrics can distinguish between algorithms that might otherwise appear similar in raw performance but differ fundamentally in coordination logic.

4. Results and Discussion

The study revealed deep insights into the behavior of the tested algorithms:

• Performance vs. Coordination: While DRL and Greedy achieved similar high levels of task completion (PTA), the metrics revealed different underlying behaviors. DRL showed superior environmental modeling (lower RMSE) and better workload fairness (lower Gini Index) due to learned coordination.
• Sensitivity and Robustness: The CSSC metric showed that DRL is more robust to "noisy" partners than the Greedy approach. Greedy's performance collapses more abruptly when scouts become unreliable, whereas DRL's learned policy allows for better adaptation to imperfect information.
• The Role of Scouts: Through Marginal Contribution (MC) analysis, the authors proved that scouts are the "critical facilitators." Removing a scout had a significantly higher negative impact on both collection (PTAC) and model accuracy (RMSE) than removing a forager, highlighting the asymmetric importance of information acquisition in sequential tasks.
• Lévy Walks Limitations: While Lévy Walks provided broad coverage, they showed high Discovery-to-Service Latency (DSL) and low CSR, indicating that their "cooperation" is weak because foragers often act independently of scout discoveries.

5. Significance

This work shifts the paradigm of multi-agent evaluation from "What was achieved?" to "How was it achieved?" By providing tools to measure fairness, redundancy, dependency, and robustness, the authors offer a framework that is highly applicable to real-world robotic deployments (like search-and-rescue or environmental monitoring) where understanding the reliability and coordination of heterogeneous teams is as important as the task outcome itself.