Introduction
Building an AI-powered appointment manager presents significant computational challenges, particularly when dealing with the "messiness" of real-world scheduling. Traditional temporal reasoning models, such as Allen’s Interval Algebra, often struggle with scalability and the representation of recurring events. This paper, presented at AAAI-91, addresses these gaps by extending the Simple Temporal Problem (STP) framework. The authors introduce a formal language for repeated activities and utilize calendar dates as hierarchical partitioning tools to drastically accelerate constraint checking.
Methodology
The proposed approach builds upon the STP model, where time is represented as a set of points and constraints are expressed as distances between them.
1. Representation of Repeated Activities (RA)
To avoid the network bloat caused by creating individual nodes for every instance of a recurring event (e.g., a weekly class), the authors define a syntax for periodic events:
seq: Defines a sequence, such as "all Mondays."tie: Links an activity to a sequence, such as "a 1-hour class every Monday at 4:00."
This implicit representation allows for efficient overlap detection. Using modular arithmetic and "convexification"—treating a sequence of events as a single large interval for initial filtering—the system can detect potential conflicts in near-constant time $O(1)$.
2. Dates as Reference Intervals
The authors leverage the natural hierarchy of dates (years, months, days) to partition the constraint network. Instead of a global network where every new appointment triggers a full propagation, the system uses dates as reference intervals. Constraints are primarily propagated among activities sharing the same "reference date," effectively pruning the search space and avoiding unnecessary re-calculations between distant events.
Experimental Evaluation
The methodology was evaluated using the CLASM (CLASSIC Schedule Manager) testbed. The researchers conducted experiments using simulated scheduling scenarios to compare four versions of the system:
- NC-NR: No conversion of qualitative relations to metric constraints, no reference intervals.
- C-NR: Conversion to metric constraints, no reference intervals.
- NC-R: No conversion, with reference intervals.
- C-R: Both conversion and reference intervals.
Results
The experimental results demonstrated a massive leap in efficiency when both optimization techniques were applied.
As shown in the data, the combination of metric conversion and reference intervals (C-R) reduced the average time per assertion to approximately 5% of the baseline (NC-NR), dropping from 97.89 seconds to just 5.0 seconds. Furthermore, storage efficiency improved significantly, with the number of stored constraints reduced to roughly 28% of the original network size. The use of dates as pruning tools allowed the system to maintain a manageable network size even as the number of appointments grew.
Conclusion
By treating dates as "transparent" reference objects and using mathematical shortcuts for repeated activities, the authors successfully bridged the gap between theoretical temporal reasoning and practical scheduling. While the current model has limitations—such as a restricted language for complex recurring patterns (e.g., "first Friday of the month") and difficulties with disjunctive constraints—it provides a robust foundation for efficient, real-time personal calendar management. Future work may extend these partitioning techniques to other fixed reference events beyond the standard calendar.
Summary updated