Introduction
In the mid-1980s, artificial intelligence research faced a significant hurdle in modeling dynamic, multiagent environments. Dominant frameworks like STRIPS and Situation Calculus were primarily designed for single agents acting in static worlds. These systems relied heavily on "frame axioms"—explicit statements defining what remains unchanged when an action occurs. However, as world complexity increases, the number of these axioms grows exponentially, leading to the notorious "Frame Problem." Furthermore, these models struggled with simultaneity; if two agents attempted to change different variables at once, their individual "stay the same" requirements often created logical conflicts.
A Model-Theoretic Approach to Events
Michael Georgeff’s 1986 paper introduces a transition-based model to overcome these limitations. Instead of defining actions by what they preserve, Georgeff defines an event as a set of possible state sequences. This shift allows for a more permissive representation of actions.
In this framework, an action is described by the constraints it imposes on the world rather than a complete list of its effects. For example, an action might specify that "Variable A must become 1," while leaving "Variable B" free to vary. This "freedom to vary" is crucial for multiagent domains, as it allows other agents or environmental processes to modify unrelated properties simultaneously without causing logical contradictions.
The Law of Persistence and Concurrency
A key innovation in Georgeff’s model is the decoupling of persistence from individual action definitions. He proposes a global Law of Persistence: if no occurring action directly affects a specific property, that property persists into the next state. This is formalized through "Direct-Effects Formulae" ($bp$), which identify the parts of the world an action intentionally modifies.
To handle simultaneous actions, the model uses the Intersection of Behaviors. The result of two concurrent actions is simply the intersection of their possible transitions. If Agent 1 moves Block A and Agent 2 moves Block B, the valid resulting state is the one where both constraints are satisfied. This approach avoids the "interleaving approximation"—the common practice of treating simultaneous events as if they happened sequentially.
Handling Complex Dependencies
The paper also addresses "contingent facts" through Causal Laws. In scenarios like a seesaw, moving one end physically necessitates the movement of the other. Georgeff’s model allows the primary action (moving one end) to trigger a causal axiom that moves the other end simultaneously. This separates the simple intent of the agent from the complex physical logic of the environment.
Conclusion and Future Directions
Georgeff’s framework provides a robust semantics for concurrent events while mitigating the Frame Problem. By moving persistence to a global level and making action definitions permissive, the model achieves greater tractability than traditional frame-axiom-based systems. However, challenges remain, particularly regarding the undecidability of consistency checking in non-propositional cases and the difficulty of reasoning under incomplete information about other agents' actions. Future work in this area continues to explore how default rules and circumscription can further refine these multiagent models.
Summary updated