TLDR: This research explores “legibility” in AI, which is how clearly an agent communicates its intentions through its actions. It extends this concept to multi-agent teams, proposing a method for AI agents to act more legibly. Experiments show that teams with a legible leader complete collaborative tasks faster and allow teammates to understand intentions quicker, especially in purely cooperative scenarios. In competitive settings, the benefits are present but less pronounced in larger environments, suggesting a need for balancing legibility with optimal behavior.
The paper “Teammates, Am I Clear?”: Analysing Legible Behaviours in Teams explores how artificial intelligence agents can communicate their intentions more clearly to their teammates, especially in collaborative tasks. The concept of “legibility” in AI refers to how easily an observer can understand an agent’s goals from its actions. Traditionally, legibility has been studied in scenarios where a single agent interacts with a human. This research extends this idea to multi-agent settings, where teams of AI agents work together.
The core problem addressed is that while optimal AI agents are efficient, their actions might not always be easy for other agents (or humans) to interpret. Imagine building a cabinet with a partner. If you first gather all the pieces for the frame before starting assembly, it clearly signals your intention to build the frame first, allowing your partner to focus on drawers. This is an example of legible behavior. The researchers investigate whether using such legible decisions can improve the efficiency of a team in collaborative tasks.
To achieve this, the paper proposes an extension of the Policy Legible MDP (PoLMDP) framework for multi-agent cooperative scenarios. This approach uses a “centralized training and decentralized execution” model. During training, each agent learns its own decision policy, considering the actions of other agents. At execution, agents act independently. The training process involves first learning an optimal joint policy for all agents, and then, for designated “legible agents,” learning a legible policy that is a “best response” to the actions of their optimal teammates. This allows for flexible adaptation and independent agency while still promoting clear communication.
The study tested its hypotheses using two common multi-agent reinforcement learning scenarios: the Level-Based Foraging (LB-Foraging) benchmark and the Pursuit-Evasion problem. In LB-Foraging, teams of agents must collaborate to pick up food items, requiring the sum of their levels to meet or exceed the food item’s level. The experiments focused on scenarios where agents had to work together for every item. In the Pursuit-Evasion scenario, a team of “hunters” must capture “preys” before time runs out, with preys trying to evade.
The teams in the experiments consisted of a “leader” agent and a “follower” agent. The leader was responsible for implicitly communicating the sequence of objectives through its actions. The follower agent had to infer these objectives. Two conditions were compared: an “optimal condition” where the leader followed a standard optimal policy, and a “legible condition” where the leader followed a PoLMDP policy. The follower agent always used an optimal policy.
For the LB-Foraging scenario, the results strongly supported the hypotheses. Teams with a legible leader consistently outperformed teams with an optimal leader, requiring significantly fewer steps to complete tasks. Furthermore, the follower agents in legible teams were able to infer the leader’s next objective much faster. This improvement was even more pronounced in larger grid sizes, suggesting that legible agents can leverage greater flexibility in less constrained environments to provide more informative behavior, enhancing team efficiency.
In the Pursuit-Evasion experiments, the benefits of legible behaviors were more nuanced. While legible teams performed better or comparably to optimal teams, especially in smaller grid sizes, the advantage diminished in larger grids. The follower agents in legible teams still inferred objectives faster in smaller to medium grids, but this difference became less significant in the largest grids. The researchers suggest this might be due to the competitive nature of the task; in larger spaces, optimal leaders’ direct actions might be as effective as legible leaders’ signaling, as preys have more room to maneuver.
Overall, the research concludes that legible behaviors have a positive impact on teamwork, particularly in purely collaborative settings where objectives are aligned. Legible agents are better at conveying their intentions, which is crucial for applications like ad-hoc teamwork where new members need to quickly understand team goals. However, in competitive scenarios, a balance between optimal and legible behaviors might be necessary, as excessive transparency could potentially be exploited by opposing teams. Future work could explore mixed strategies where agents can switch between optimal and legible behaviors as needed.
Also Read:
- Unveiling Intentions: How AI Agents Actively Infer Goals in Complex Environments
- Unlocking AI Teamwork: A New Framework for Interpretable Multi-Agent Learning
For more details, you can refer to the full research paper available at this link.
