Abstract
Automated travel planning often suffers from a data gap between user preferences and actionable schedules. The OnDroad Planner addresses this by integrating crowdsourced data from the MINUBE social network into a formal planning framework. By combining geographic clustering with PDDL-based optimization via the Metric-FF planner, the system generates personalized multi-day itineraries that respect temporal constraints and social popularity. This approach transforms social "likes" into structured, executable plans while accounting for real-world constraints like transit times and opening hours.
Introduction
In the modern tourism industry, automated planning systems are limited by the quality of their input data. Defining the "initial state" (a user's location and available resources) and "goals" (desired attractions) is difficult because information is often fragmented across various services. While platforms like TripAdvisor or MINUBE provide excellent recommendations through collaborative filtering, they do not compose a plan. Users are typically left to manually arrange these suggestions into a logical schedule.
The OnDroad Planner aims to bridge this gap by automating the entire pipeline: from social network recommendations to a fully scheduled itinerary. By leveraging the social graph and geographic heuristics, the system replicates human-like planning logic within a formal automated planning environment.
System Architecture
The ONDROAD system architecture consists of three primary stages designed to move from unstructured social data to a structured PDDL (Planning Domain Definition Language) problem.
1. Information Extraction
The system integrates with the MINUBE social network to query the "N best" Points of Interest (POIs) in a destination. These POIs are filtered based on popularity within the user’s social circle. This effectively turns social "votes" into planning goals, ensuring the generated plan reflects both general popularity and personal social relevance.
2. Geographic Clustering
To mimic human intuition—where travelers explore one neighborhood at a time to minimize transit fatigue—the system employs k-means clustering. POIs are grouped geographically, and each cluster is assigned to a specific day of the trip. The Tourist Plan Manager then calculates estimated travel times (walking or driving) between these points and integrates POI timetables to ensure feasibility.
3. Automated Planning with Metric-FF
The core of the system is a planning engine that formulates the itinerary as a PDDL problem.
- Actions: The domain includes visiting a POI, walking, driving, eating, and sleeping.
- Soft Goals: Recognizing that time is limited, the planner treats POIs as "soft goals." Each goal is assigned a preference value derived from social network ratings, allowing the planner to prioritize high-value attractions.
- Optimization: The system utilizes the Metric-FF planner, which is capable of handling numeric preconditions such as time and schedules. This allows the system to maximize the total "value" of the trip within the user's temporal constraints.
Data and Implementation
The primary data source is the MINUBE social network, leveraging its database of POIs, user-generated photos, geo-localization data, and recommendation "votes." Unlike traditional pathfinding tools, the OnDroad Planner treats travel as a complex planning task. It includes "generic actions" like eating and sleeping, which are essential for a realistic tourist experience but often overlooked in classical algorithms.
One of the system's most innovative features is the concept of "Human-in-the-Loop Data." The authors suggest that users can act as sensors; if a user updates a museum's opening hours within the app, the system can trigger a re-planning service for all affected users, using human computation to maintain a dynamic and accurate database.
Results and Discussion
As a work-in-progress, the project has successfully demonstrated a functional pipeline that moves from unstructured social data to a structured PDDL problem. Key achievements include:
- System Integration: Successful coupling of social media APIs with formal planning engines.
- Heuristic Mimicry: The use of k-means clustering effectively replicates the way humans naturally group activities to reduce transit overhead.
- Planner Selection: Metric-FF was identified as a robust engine for handling the numeric constraints necessary for realistic scheduling.
Limitations and Future Work
Several challenges remain for the full deployment of the OnDroad Planner. Dynamic Re-planning is currently under development to handle real-world delays, such as a lunch that runs over schedule. Additionally, the system's efficacy is tied to the density of the MINUBE network; in regions with sparse data, the quality of generated plans may decrease. Future iterations aim to incorporate more complex user constraints, such as budget limits and specific dietary preferences for the "eat-at" actions, further personalizing the automated travel experience.
Summary updated