Abstract
This paper investigates how artificial agents transition from deliberate planning to routine behavior by grounding their actions in specific environments, or "habitats." The authors argue that routine behavior is not merely an internal mental shortcut but a result of an agent's historical "coupling" with its environment. Using the FLOABN system, the study demonstrates that familiarity with a specific habitat significantly reduces the computational and perceptual workload required to achieve goals.
Introduction
In classical AI, agents often rely on heavy internal representations and planning from first principles. However, humans rarely plan familiar tasks, such as making coffee, from scratch; they rely on routines. This research addresses the gap between reactive agents and planning agents by exploring how an agent’s history with a specific physical location—a habitat—allows it to focus only on the features necessary for a task while ignoring irrelevant details. The core challenge is quantifying how this "environmental fit" saves effort in terms of perception and information processing.
The FLOABN System and Multicase Memory
The researchers utilize the FLOABN system, which centers on a unique data structure called the Multicase. Unlike a static script, a multicase is a directed graph of Decision Points (DPs). Each DP contains "descriptors" (environmental features) and "options" (actions).
The system employs SCAVENGER, an adaptive planner that retrieves previous cases from memory and adapts them to current situations. As the agent interacts with the world, it "enriches" the multicase by adding new decision points or refining existing ones, effectively using the environment as an external memory anchor.
Measuring Agent Efficiency
To measure efficiency, the authors developed a weighted workload metric:
$$Workload = Steps + FeaturesAttended + 2(FeaturesStored) + 5(ActiveLooks) + 10(Adaptations)$$
This formula penalizes "expensive" cognitive tasks, such as state-space searching (adaptations) and visual searching (active looks), providing a quantitative look at how routines reduce cognitive load.
Experimental Results
The study modeled 50 episodes across three domains: telephoning, photocopying, and vending machines. The sequence included repetitions and variants to simulate the development of a routine. The empirical tests showed a clear "power law" style improvement in efficiency as episodes progressed.
As shown in the workload analysis, the slope of the graph flattens over time. Once a routine is established, the agent "knows" what to ignore, virtually eliminating the need to search for specific objects (like power switches or coin slots) in familiar habitats. The marginal effort required to complete tasks reaches a low, stable baseline.
Conclusion and Future Work
The "Habitat Advantage" suggests that performance gains in AI come not just from internal knowledge transformation, but from improving the fit between the agent and its world. While the current results are based on discrete-event simulators, they provide a foundation for understanding how agents can use "common practices" to interpret local variations. Future work will address the scaling of multicase structures to thousands of routines and the potential for social learning, where agents copy the actions of others within a shared habitat.
Summary updated