Learning from Evolution: Improving Collective Decision-Making Mechanisms using Insights from Evolutionary Robotics

TL;DR

This work addresses fast, accurate collective decision-making in multi-robot teams under incomplete information by leveraging evolutionary computation to derive efficient neural-network-based mechanisms. It analyzes evolved decision-making networks, applies SHAP-based interpretability to extract actionable insights, and hand-codes two interpretable decision rules (HC1 and HC2) that outperform voter model and majority rule in benchmarks. The results show that the hand-coded mechanisms offer a favorable mix of speed, accuracy, and interpretability, highlighting the value of combining evolutionary insights with explainable AI to produce robust, transparent control policies. The approach demonstrates a practical path from opaque evolved strategies to efficient and understandable hand-coded mechanisms suitable for scalable multi-robot systems.

Abstract

Collective decision-making enables multi-robot systems to act autonomously in real-world environments. Existing collective decision-making mechanisms suffer from the so-called speed versus accuracy trade-off or rely on high complexity, e.g., by including global communication. Recent work has shown that more efficient collective decision-making mechanisms based on artificial neural networks can be generated using methods from evolutionary computation. A major drawback of these decision-making neural networks is their limited interpretability. Analyzing evolved decision-making mechanisms can help us improve the efficiency of hand-coded decision-making mechanisms while maintaining a higher interpretability. In this paper, we analyze evolved collective decision-making mechanisms in detail and hand-code two new decision-making mechanisms based on the insights gained. In benchmark experiments, we show that the newly implemented collective decision-making mechanisms are more efficient than the state-of-the-art collective decision-making mechanisms voter model and majority rule.

Figures (7)

• Figure 1: Collective perception scenario in the BeeGround simulator. The arena is bounded by walls and the environmental features are represented by the black and white pattern on the arena floor. Robots indicate their current opinion via an LED on top (red stands for Black, blue for White).
• Figure 2: Probabilistic finite state machine for decision-making (a) and finite state machine for robot motion (b) based on Valentini et al. Valentini2016. Time periods $t^{exp}_n$, $t^{dis}_n$, $t^{rot}_n$, $t^{str}_n$ and angle $\beta_n$ are randomly sampled. Buffer values $\zeta_n < 0$ indicate that the robot is potentially stuck between obstacles.
• Figure 3: Topology of the evolved decision-making mechanisms. Inputs are the percentage of neighbors with opinion White$w(t)$, the normalized length of the message queue of neighbor opinions $l(t)$, the ground sensor value $g(t)$ at current time step $t$, and the robot's previous opinion $o(t-1)$. The ANN outputs the robot's current opinion $o(t)$.
• Figure 4: Mean consensus time $\overline{T}_N$ and exit probability $E_N$ for each of the ten best-evolved individuals in problem difficulties $\rho^* \in \{0.67, 0.82\}$. We split up the data between White-dominant (gray bars, left) and Black-dominant (black bars, right) environments. Red dashed lines indicate the best performance of our baselines voter model and majority rule, i.e., the speed (i.e., lowest $\overline{T}_N$) of the majority rule and the accuracy (i.e., highest $E_N$) of the voter model. Plots for $\rho^* \in \{0.25, 0.52\}$ are available on Zenodo zenodo_2024.
• Figure 5: Mean SHAP values for four representative runs where $w(t)$ is the percentage of neighbors with opinion White, $l(t)$ the normalized length of the message queue of neighbor opinions, $g(t)$ the ground sensor value, and $o(t-1)$ the robot's previous opinion. $w(t)$ and $l(t)$ are grouped as their values are highly correlated. Plots for all other runs and beeswarm plots are available on Zenodo zenodo_2024.