Introduction
Autonomous agents often face a trade-off between reactive behavior and long-term planning. Traditional Belief-Desire-Intention (BDI) architectures can suffer from "narrowness-boundedness," where an agent becomes so focused on a specific plan that it fails to anticipate environmental changes. This paper proposes using "surprise"—the mismatch between expectation and perception—as a functional signal to trigger shifts in an agent's mental state, allowing for efficient anticipation without exhaustive re-computation.
The Architecture of Anticipation
The proposed architecture treats surprise as a shortcut for strategy adaptation. Instead of fine-tuning individual parameters, the agent transitions between distinct "mental states" or moods based on a history of surprises. This allows for a global reallocation of resources, such as trading off movement speed for increased sensor vigilance when the environment becomes risky.
Mental State Clustering
The agent's behavior is clustered into four primary states managed by a transition logic that responds to positive, negative, or null surprise signals:
- Default: Standard exploitation of known resources.
- Excitement: Triggered by positive surprises; it reinforces successful paths and increases exploration.
- Caution: Triggered by negative surprises (e.g., detecting unexpected smoke); it prioritizes safety by increasing sensor sampling rates and reducing speed.
- Curiosity/Boredom: Triggered by a lack of environmental signals; it drives the agent to explore unvisited areas to maintain an accurate world model.
These transitions are managed by a Push Down Automaton (PDA) that tracks the history of observations, allowing the agent to respond to patterns of surprise rather than isolated, noisy events.
Experimental Evaluation: Foraging in a Risky World
The researchers tested this approach in a 2D simulation where an agent must collect food while avoiding dynamic fire hazards. Fires are preceded by smoke, providing a low-level signal that something in the environment is changing. The agent's goal is to maximize food collection while managing a finite energy supply. The Mental State (MS) agent was compared against a standard Subjective Expected Utility (SEU) agent that relies purely on utility calculations.
Key Results
In high-risk scenarios with multiple active fires, the MS agent significantly outperformed the SEU baseline. While the SEU agent often prioritized speed and collided with hazards, the MS agent's transition to a "Caution" state allowed it to navigate safely.
As shown in Figure 3, the MS agent maintains higher effectiveness as environmental riskiness increases. Although the SEU agent might be more efficient in low-risk, short-term tasks, the MS agent's "epistemic investment" through Curiosity ensures better long-term performance by preventing the agent's internal world model from becoming obsolete.
Conclusion
By using surprise as a functional shortcut, agents can achieve a form of "bounded anticipation." This mechanism allows them to reallocate resources globally without the computational overhead of continuous replanning. Future work will explore these mental states in social contexts, where surprise might signal changes in trust or competitive dynamics between multiple agents.
Summary updated