Abstract
Humans have developed jurisprudence as a mechanism to solve conflictive situations by using past experiences. Following this principle, we propose an approach to enhance a multi-agent system by adding an authority which is able to generate new regulations whenever conflicts arise. Regulations are generated by learning from previous similar situations, using a machine learning technique (based on Case-Based Reasoning) that solves new problems using previous experiences. This approach requires: to be able to gather and evaluate experiences; and to be described in such a way that similar social situations require similar regulations. As a scenario to evaluate our proposal, we use a simplified version of a traffic scenario, where agents are traveling cars. Our goals are to avoid collisions between cars and to avoid heavy traffic. These situations, when happen, lead to the synthesis of new regulations. At each simulation step, applicable regulations are evaluated in terms of their effectiveness and necessity. Overtime the system generates a set of regulations that, if followed, improve system performance (i.e. goal achievement).
Introduction
In Multi-Agent Systems (MAS), autonomous agents often interact in ways that lead to conflicts, such as collisions in robot swarms or congestion in networks. Traditionally, these issues are addressed by hardcoding "social laws" at design-time or using exhaustive state-space searches. However, hardcoding is brittle and cannot handle unforeseen situations, while exhaustive search is often computationally prohibitive. This research addresses the challenge of on-the-fly norm synthesis: enabling a system to observe a new conflict, propose a regulation, and determine its quality without expert guidance.
Methodology
The core of the proposed approach is an unsupervised Case-Based Reasoning (CBR) cycle managed by a central "regulatory authority." This authority monitors the scenario for conflicts and manages the lifecycle of norms.
The process follows these steps:
- Observation: Monitor agents for conflicts.
- Retrieve: When a conflict occurs, the system searches for similar past situations in its case base.
- Reuse/Generate: If a similar case exists, the system adapts the previous solution. Otherwise, it generates a new solution (e.g., "Car A must stop").
- Translation: The solution is formalized into a norm:
IF (condition) THEN obl(action). - Revise & Evaluate: The authority tracks norm compliance and its impact on system goals (safety and efficiency).
- Retain: The case base is updated with the norm's performance score.
The system relies on the Continuity Assumption, which posits that similar social situations require similar regulations. This allows the CBR to generalize from a limited number of cases to a wide variety of encounters. Norms are not permanent; if a norm's effectiveness or necessity score falls below a threshold, it is deactivated.
Experimental Setup
The approach was evaluated using a custom-built traffic simulation in the Repast Simphony framework. The environment consisted of a 20x20 orthogonal road junction. Simulations ran for 10,000 ticks, with new cars added every 2 ticks. To test robustness, agents had a 30% probability of violating norms.
Results and Discussion
Two primary experiments were conducted:
- Safety Only: When the primary goal was collision avoidance, the system rapidly converged to a stable set of approximately 10 norms. Collisions dropped to zero after 550 simulation ticks.
- Safety vs. Fluidity: When minimizing traffic stops was also a goal, the system faced a trade-off. It dynamically toggled norms to balance safety and speed, significantly reducing collisions compared to an unregulated scenario while maintaining better traffic flow than the safety-only configuration.
Conclusion
This framework demonstrates that artificial societies can "learn" their own laws through experience, mimicking human jurisprudence. While the current traffic model is simplified, the results show that combining top-down regulation with bottom-up feedback creates an effective closed-loop learning system. Future work will explore scalability in more complex environments and the interactions between different norms to ensure a consistent legal code.
Summary updated