🤖 machine learning

From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework for Personalized On-Device Itinerary Generation

The paper introduces the Plan, Learn, Adapt (PLA) framework, a three-stage on-device system that combines an ensemble of lightweight planners with a compact reward model to generate feasible and highly desirable personalized itineraries, outperforming both single planners and frontier LLMs in human evaluations while achieving production-grade latency and feasibility.

Original authors: Himel Dev, Tanmoy Sen, Madhusudan Basak, Bashima Islam

Published 2026-07-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Himel Dev, Tanmoy Sen, Madhusudan Basak, Bashima Islam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a spaceship, but instead of navigating through stars, you are navigating a bustling city to plan the perfect vacation. This is the world of trip planning, a puzzle that sits at the crossroads of two very different scientific worlds. On one side, you have combinatorial optimization, which is like a strict math teacher making sure you don't break the rules: the museum closes at 5 PM, the train leaves at 6, and you can't visit the same place twice. On the other side, you have preference learning, which is like a psychic trying to guess what you actually want to feel—do you want a relaxed day or a packed adventure? The tricky part is that these two worlds often fight each other. A mathematically perfect schedule might feel boring, and a dreamy, fun schedule might be impossible to actually execute. The big question for anyone who has ever tried to plan a trip is: How do we build a travel guide that is both legally possible to follow and genuinely delightful to experience, all while running on your phone without needing the internet?

This paper introduces a clever new system called PLA (Plan, Learn, Adapt) that tries to solve this exact problem. The authors, working with a travel app called FlyEnJoy, realized that existing methods were failing in two specific ways. Some old-school planners were great at following the rules but terrible at understanding human taste, often creating itineraries that felt robotic. Meanwhile, the newest, flashiest AI tools (like the giant language models everyone is talking about) were great at sounding helpful but completely failed at the basic math of travel; in their tests, these AI models produced itineraries that were strictly infeasible (0% success rate), with a statistical upper bound suggesting they might have been valid less than 4% of the time, violating opening hours and travel times.

To fix this, the team built a three-step "factory" that runs entirely on your phone:

Plan (The Idea Factory): Instead of relying on just one way to make a schedule, they built a team of five different "planners" (like a greedy chef, a careful accountant, and a creative artist). Each one tries to build a valid trip from scratch. Because they think differently, they come up with very different lists of places to visit, ensuring they have a diverse pool of "feasible" (rule-abiding) options to choose from.
Learn (The Taste Tester): The team didn't just guess what people liked; they asked humans to compare pairs of full-day itineraries. They collected over 2,500 of these comparisons. Using this data, they trained a tiny, super-fast "reward model" (a digital taste-tester) that learned to spot the invisible qualities of a great trip, like good pacing and a nice balance of activities, rather than just counting how many museums were visited.
Adapt (The Local Refiner): Once the best itinerary is picked, this stage makes small, safe tweaks to it. It's like a chef tasting a soup and adding a pinch of salt, but with a strict rule: every single change must keep the recipe valid. If a tweak breaks the schedule (like making you arrive at a closed shop), it is instantly thrown away. This happens so fast that it takes only 109.9 milliseconds on average, meaning you get a perfect, personalized plan instantly, even if you are on a plane with no Wi-Fi, as the system is designed for offline-first use.

The results are impressive. When they tested their system against the best single planner, the "Plan, Learn, Adapt" team won 67.8% of the time against human judges. In contrast, when they asked three of the most advanced AI models (GPT-5, Claude Opus 4.5, and Gemini 3 Pro) to create itineraries under the same strict rules, none of them (0%) produced a single valid trip. The new system also generalized well, working accurately in over 100 different U.S. cities it had never seen before.

In the real world, when they deployed this system into the FlyEnJoy app, the results were tangible. Users started their trip planning 83% more often, and 91% more people actually finished their itineraries. The system didn't just make the plans; it made the whole experience smoother, proving that you can have a travel guide that is both mathematically perfect and perfectly fun, all while living right inside your pocket.

Technical Summary: From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework

1. Problem Statement

The paper addresses the challenge of generating personalized, multi-day travel itineraries on mobile devices. This task is defined by a fundamental tension between hard combinatorial feasibility and soft latent desirability:

Feasibility: Itineraries must strictly adhere to spatio-temporal constraints, including POI opening hours, travel times, daily time budgets, and cost limits.
Desirability: Itineraries must satisfy subjective user preferences regarding pacing, geographic coherence, and category diversity, which are difficult to capture with static, hand-crafted objective functions.
Viability: Solutions must be deployable on end-user devices (specifically offline-first iOS environments) with strict latency requirements (interactive speed) and privacy constraints, ruling out cloud-dependent Large Language Models (LLMs).

Existing approaches fail to meet all three requirements simultaneously: classical optimizers ensure feasibility but lack nuanced preference modeling; learning-based methods capture preferences but often violate hard constraints; and LLMs, while capable of natural language generation, frequently produce invalid schedules (0% feasibility in the authors' benchmark) and are incompatible with offline, low-latency mobile deployment.

2. Methodology: The PLA Framework

The authors propose Plan, Learn, Adapt (PLA), a three-stage framework designed to jointly address feasibility, desirability, and viability.

Phase 1: PLAN (Diverse Feasible Generation)

The Plan stage generates a structurally diverse set of feasible candidate itineraries using a heterogeneous ensemble of five lightweight planning algorithms, all sharing a unified feasibility layer:

Algorithms: Greedy (constructive), Dynamic Programming (DP), Beam Search, A* Search, and Simulated Annealing (SA).
Feasibility Layer: A shared logic module that enforces hard constraints (visit hours, travel time consistency, time/cost budgets, no repeats) and computes earliest feasible arrival times.
Objective: While all planners optimize a shared composite score (balancing travel efficiency, waiting time, urgency, category diversity, and POI popularity), their distinct search paradigms yield structurally different itineraries. This diversity is crucial because no single algorithm dominates across all trip contexts.

Phase 2: LEARN (Human Preference Modeling)

The Learn stage fits a compact reward model to capture latent user satisfaction at the itinerary level, rather than relying on per-POI scores.

Data Collection: 2,519 pairwise human comparisons of complete itineraries across 100+ U.S. cities were collected. Annotators chose between itineraries generated by different planners without knowing the source.
Model Architecture: A Bradley-Terry reward model is trained on ~20 interpretable itinerary-level features (e.g., total travel time, category entropy, daily load balance, schedule slack).
- The model uses a difference-based formulation: $P(x_a \succ x_b) = \sigma(R_\theta(x_a) - R_\theta(x_b))$ .
- Monotonicity Constraints: The model enforces domain knowledge (e.g., increased travel time should not increase reward) directly during optimization.
Selection: The reward model acts as a selector, choosing the highest-scoring itinerary from the diverse pool generated by the Plan stage.

Phase 3: ADAPT (On-Device Refinement)

The Adapt stage performs reward-guided local refinement on the selected seed itinerary within a strict device-aware compute budget.

Feasibility-Preserving Edits: The system applies atomic operators (Drop, Insert, Replace, Swap, Relocate, MoveDay) to the itinerary. Crucially, every edit is followed by an incremental re-timing procedure that validates constraints. If an edit violates a constraint, it is discarded. This guarantees that every intermediate state remains feasible, unlike post-hoc repair methods.
Search Strategy: A hill-climbing approach with stochastic neighborhood sampling selects moves that improve the reward score.
Resource Awareness: The refinement process is bounded by a dynamic time budget (typically ~200–500 ms) adjusted based on device signals (core count, memory, power mode) to ensure interactive latency.

3. Key Contributions

Heterogeneous Planner Ensemble: The authors empirically demonstrate that no single planning algorithm dominates. By combining five distinct algorithms, they achieve a 67.8% win rate in pairwise comparisons, an 11.2 percentage point improvement over the best single planner (DP).
Itinerary-Level Preference Modeling: Moving beyond per-POI scoring, the paper introduces a compact Bradley-Terry model trained on complete itinerary comparisons. This model captures emergent properties (pacing, day balance) and achieves 67.6% accuracy in leave-one-city-out cross-validation.
On-Device Feasibility-Preserving Refinement: The Adapt stage improves desirability while guaranteeing 100% feasibility by construction. It achieves an average on-device latency of 109.9 ms, enabling fully offline, interactive execution.

4. Experimental Results

Feasibility vs. LLMs: In a benchmark against three frontier LLMs (GPT-5, Claude Opus 4.5, Gemini 3 Pro) on 294 itineraries, the LLMs achieved 0% strict feasibility, violating constraints such as travel time and visit hours. In contrast, PLA maintained 100% feasibility.
Ensemble Performance: The reward-guided ensemble outperformed the best individual planner by 11.2 points (67.8% vs. 56.6% win rate). The selection distribution showed high entropy (97.6% of maximum), confirming that different algorithms contribute uniquely depending on the context.
Refinement Uplift: The Adapt stage provided statistically significant score improvements across all seed algorithms. For example, A* seeds saw an 11.7% improvement, while the already high-scoring ensemble seeds improved by 3.9%.
Production Impact: Deployed in the FlyEnJoy iOS app, PLA increased itinerary completion rates by 91% and initiations by 83% compared to a single-heuristic baseline, while reducing average session time per user by 38%.

5. Significance and Claims

The paper claims that PLA provides a viable blueprint for human-aligned planning under real-world deployment constraints. Its significance lies in resolving the trade-off between feasibility and desirability without relying on cloud-based AI:

Robustness: By separating feasibility (Plan) from preference learning (Learn) and using feasibility-preserving refinement (Adapt), the system avoids the "invalid schedule" problem common in learning-based and generative AI approaches.
Portability: The use of a compact, interpretable reward model and lightweight search operators allows the system to run entirely on-device, supporting offline use cases (airports, in-flight) where cloud APIs are unavailable or too costly.
Practicality: The framework demonstrates that complex, multi-day planning can be executed with interactive latency (<300 ms) while strictly adhering to hard constraints, a capability that current frontier LLMs lack in this specific domain.

The authors conclude that while LLMs struggle with hard constraints, a structured, hybrid approach combining classical optimization, preference learning, and constrained local search can deliver superior, production-ready results for personalized itinerary generation.