Introduction
In the early development of artificial intelligence, solving a single, well-defined problem was a significant milestone. However, real-world environments often require an agent to satisfy multiple, non-independent goals simultaneously. Traditional AI systems typically addressed this by attempting to synthesize a single, rigid "global plan" before execution. This approach is inherently brittle; any environmental change or the addition of a new goal often necessitates a complete and computationally expensive recalculation of the entire plan.
In this paper, Arthur M. Farley proposes a shift from static global planning to dynamic execution coordination. The central question is how a system can manage multiple independent strategies concurrently, making real-time decisions to ensure that progress toward one goal does not permanently obstruct the path to another.
The Coordination Framework
Farley introduces a model where strategies are represented as trees, constructed "backward" from the desired goal state using a breadth-first search. To facilitate dynamic decision-making, the system utilizes a specific state decomposition:
- SSN (Static/Non-changeable): Propositions that the system cannot influence (e.g., environmental constants).
- SSO (Operator-changeable): Propositions that the system can modify through its available actions.
Dynamic Classification and Execution
Rather than executing a linear sequence of steps, the system operates in a continuous cycle of classification and action selection:
- Relevance Check: A strategy is considered "relevant" only if its unchangeable conditions (SSN) match the current state of the world.
- Realization Check: A state within a strategy is "realized" if the current situation already satisfies all its requirements.
- Grounding and Critical States: This is the core logic of the coordination process. A state is defined as critical if it is a bottleneck—meaning every possible path to the goal must pass through it. A strategy becomes ungrounded (temporarily suspended) if pursuing its goal would destroy a critical state required by another active strategy.
By identifying these bottlenecks, the system can distinguish between minor conflicts (where a state can be easily recreated) and catastrophic conflicts (where a necessary condition is permanently or difficultly lost).
Implementation in the Blocks World
The proposed model was implemented and evaluated within the "Blocks World" domain—a classic symbolic AI environment where a robotic arm must move blocks into specific configurations. While the paper does not provide the quantitative benchmarks common in modern machine learning, it demonstrates a functional logic for symbolic problem-solving. The system successfully identified when strategies interfered with one another, prioritized "grounded" strategies to ensure steady progress, and automatically resumed "ungrounded" strategies once the environmental conflicts were resolved.
Key Insights
- Postponed Coordination: By delaying the coordination of goals until the execution phase, the system remains highly flexible. New goals can be added to a "pending" set at any time without triggering a full replanning cycle.
- The Critical State Heuristic: This provides a logical basis for "why" a conflict matters, allowing the system to prioritize actions based on the necessity of certain states rather than simple proximity to a goal.
- Hierarchical Abstraction: The paper suggests that high-level strategies should only be refined into tactical details when necessary, supporting a more efficient hierarchical approach to complex problem-solving.
Limitations and Future Research
Farley identifies several areas for further development:
- Satisfiability: The system currently lacks a robust method for determining if a set of conjunctive goals is logically possible to satisfy simultaneously before attempting them.
- Resource Constraints: The model focuses primarily on logical state conflicts but does not deeply address physical resource limitations (e.g., two goals requiring the same tool at the same time).
- Computational Complexity: As the number of goals and the depth of strategy trees increase, the overhead of the cyclic classification process may become a performance bottleneck.
Conclusion
Farley’s work offers a foundational perspective on multi-goal management, moving away from the rigidity of early global planners toward a more adaptive, execution-oriented architecture. By focusing on critical states and dynamic grounding, the system maintains progress in complex, changing environments.
Summary updated