Abstract
In the field of autonomous agents, a persistent tension exists between situated activity—where agents react immediately to environmental cues—and memory-based planning, where agents follow stored sequences to achieve long-term goals. This paper introduces MAYOR, a system designed to bridge this gap within the dynamic environment of SimCity. By utilizing a network of specialized "advocates" inspired by Marvin Minsky’s Society of Mind, MAYOR manages the "Problem of Interruption." The system demonstrates that robust planning in unpredictable worlds is less about exhaustive search and more about a sophisticated theory of attention that allows an agent to arbitrate between competing long-term achievement goals and immediate maintenance tasks.
Introduction
Artificial Intelligence research has historically been divided between two paradigms. Situated agents are highly responsive to their surroundings, making them excellent at survival in chaotic environments but poor at executing complex, multi-step infrastructure projects. Conversely, memory-based planners can execute sophisticated strategies by retrieving stored plans, yet they often prove brittle when faced with unexpected environmental shifts.
The MAYOR project seeks a middle ground. It addresses the fundamental question: How can an agent pursue abstract, long-term goals while remaining interruptible by urgent, short-term maintenance needs? Using the simulation game SimCity as a generative testbed, this work explores how an agent can balance the strategic (e.g., building an industrial park) with the tactical (e.g., responding to a sudden earthquake).
The MAYOR Architecture
MAYOR is built upon a modular architecture of specialized agents known as advocates. This design avoids the pitfalls of a monolithic utility function by distributing goals across a network of competing interests.
1. Advocate Model
Inspired by Minsky’s Society of Mind, the system consists of various advocates dedicated to specific domains. For instance, a "Firefighter" advocate monitors the world state for disasters, while an "Urban Planner" advocate focuses on expansion. Some advocates are meta-level, monitoring other advocates to settle resource disputes.
2. Conflict Resolution
To prevent "analysis paralysis," MAYOR employs two layers of conflict resolution:
- Compile-time Resolution: Inherently contradictory goals are handled at the structural level, where advocates for mutually exclusive outcomes are designed to inhibit one another.
- Run-time Arbitration: Since the environment is unpredictable, MAYOR uses heuristic arbitration strategies to handle interruptions during execution. A key strategy is "seize-cheap-opportunity," which allows the agent to pause a massive, resource-intensive project if a low-cost, high-benefit action (like fixing a small road gap) becomes available nearby.
3. Hierarchical Goal Management
The system distinguishes between maintenance goals (keeping the city functional) and achievement goals (expanding the city's footprint or economy). The core challenge lies in the logic used to decide which goal "gets the floor" at any given moment.
Methodology and Environment
Unlike traditional AI research that relies on static datasets, this study utilizes SimCity as a dynamic, generative environment. The "data" in this context is the evolving state of the city, including population density, pollution levels, crime rates, budget constraints, and random disasters. This environment requires the agent to process a continuous stream of information and make real-time decisions that have long-lasting consequences.
Key Insights: Planning as Attention Management
The central insight of the MAYOR project is that planning in a complex, situated world is fundamentally a problem of attention management. Survival and success depend on the agent's ability to decide what to ignore and what to allow as an interruption.
Rather than attempting to calculate an optimal path through an exhaustive search—which is computationally intractable in a simulation as complex as SimCity—MAYOR uses its advocate network to perform heuristic arbitration. This allows the agent to make "good enough" decisions rapidly, ensuring it remains both goal-oriented and environmentally aware.
Results and Discussion
Early development of MAYOR has yielded promising qualitative results. The system successfully services long-term objectives, such as maintaining a minimum municipal income, while simultaneously reacting to localized crises like crime spikes or fires.
A primary success of the model is the functional bridge between planning and reactivity: the "Urban Planner" advocate can be successfully interrupted by the "Firefighter" advocate when a crisis occurs, and then resume its long-term trajectory once the threat is neutralized. These results suggest that the success of a situated agent hinges entirely on its theory of attention—the underlying logic that governs how and when priorities shift.
Limitations and Future Work
While MAYOR demonstrates the viability of the advocate model, it currently only manages a subset of SimCity’s complexities, such as fire, crime, and basic income. Expanding the agent's scope to handle the full range of city management variables remains a significant challenge. Furthermore, while "seize-cheap-opportunity" is an effective heuristic, future work is needed to develop more nuanced arbitration strategies that can handle complex trade-offs between long-term investment and short-term stability. The ultimate goal is to refine the theory of attention to ensure agents are neither too "distracted" by trivialities nor too "stubborn" to react to catastrophes.
Summary updated