🤖 AI

Generalizable Multi-Task Learning for Wireless Networks Using Prompt Decision Transformers

This paper proposes a Prompt Decision Transformer-based multi-task learning framework that reformulates multi-cell selection in wireless networks as a sequence modeling problem, enabling robust, few-shot generalization and significant quality-of-experience improvements across diverse and dynamic network configurations without the need for retraining.

Original authors: Fatih Temiz, Shavbo Salehi, Melike Erol-Kantarci

Published 2026-06-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fatih Temiz, Shavbo Salehi, Melike Erol-Kantarci

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 a bustling city where thousands of people (users) are trying to make phone calls or stream videos simultaneously. The city is covered by a grid of cell towers (base stations). In a perfect world, every person would connect to the perfect tower at the perfect time to get the fastest, clearest signal.

However, in reality, this is a chaotic puzzle. People are moving, buildings block signals, and too many people trying to use the same tower causes a "traffic jam" of data, leading to dropped calls and buffering. This is the problem of Radio Resource Management (RRM) in wireless networks.

Here is a simple breakdown of what this paper proposes to solve that problem:

The Old Way: The "Overworked Tutor"

Traditionally, networks use rules or "Reinforcement Learning" (AI that learns by trial and error) to decide which tower connects to which person.

The Problem: Think of a traditional AI tutor who is great at teaching a class of 5 students. If you suddenly add a 6th student, or change the classroom layout, the tutor gets confused. They have to stop, forget everything they learned, and start over from scratch to learn the new setup.
The Result: This is slow, expensive, and the AI struggles to adapt when the network gets bigger or more complex.

The New Way: The "Super-Adaptable Guide" (PromptDT)

The authors propose a new AI system called Prompt Decision Transformer (PromptDT). They treat the network problem like a story or a sequence of events, similar to how a Large Language Model (like the one you are talking to now) reads a story to predict the next word.

Here is how it works using a creative analogy:

1. Learning from a "Library of Stories" (Offline Learning)

Instead of the AI wandering around the network making mistakes in real-time (which is slow and risky), it studies a massive library of past "stories" (data logs). These stories show what happened when different numbers of people used different numbers of towers under different conditions.

The Analogy: Imagine a chess grandmaster who has studied millions of past games. They don't need to play a new game to learn the rules; they just look at the history of similar games to know what move to make.

2. The "Magic Prompt" (Few-Shot Adaptation)

This is the paper's secret sauce. In the past, if you wanted the AI to handle a new city size (e.g., 10 towers instead of 5), you had to retrain the whole system.

The Analogy: With PromptDT, you don't retrain the AI. Instead, you give it a tiny "sticky note" (a prompt) at the start of the game. This note says, "Hey, today we have 10 towers and 15 people. Here is a short example of how a good game looks in this specific setup."
The Result: The AI instantly understands the new rules based on that short note and the library of stories it already read. It adapts immediately without needing to go back to school.

3. The "Multi-Task" Superpower

The researchers tested this system on 12 different "scenarios" at once—ranging from tiny networks with 5 users to huge networks with 15 users, using different scheduling rules.

The Analogy: Instead of hiring 12 different tutors for 12 different class sizes, they trained one single super-tutor who can teach all 12 classes simultaneously. When a new class size appears (one the tutor hasn't seen before), the tutor just looks at the "sticky note" prompt and figures it out.

What Did They Find?

The paper ran simulations to see how well this "Super-Adaptable Guide" performed compared to the old "Overworked Tutor" (PPO):

Better Quality of Experience (QoE): The new system improved the user experience (less buffering, faster speeds) by up to 49% in complex, crowded scenarios.
Scaling Up: As the AI model got "smarter" (larger size), it got even better at handling the chaos of big networks.
No Retraining Needed: When they tested the AI on a completely new network setup it had never seen, it performed almost as well as a system that had been specifically trained on that exact setup from scratch.
Efficiency: Despite being powerful, the system is fast enough to run on standard computer chips, making it ready for real-world use.

The Bottom Line

This paper introduces a way to make wireless networks "smarter" and more flexible. Instead of building a new AI for every new network configuration, they built one flexible AI that can read a short "instruction note" (prompt) and instantly adapt to any network size or complexity, learning from past data to make better decisions instantly.

Technical Summary: Generalizable Multi-Task Learning for Wireless Networks Using Prompt Decision Transformers

Problem Statement
Future wireless networks require rapid adaptation to heterogeneous environments and dynamic task configurations, necessitating a shift from rule-based Radio Resource Management (RRM) to Artificial Intelligence (AI)-driven approaches. A critical RRM challenge is Coordinated Multipoint (CoMP) transmission, specifically the multi-cell selection problem. This involves determining optimal serving-cell combinations for User Equipments (UEs) to mitigate inter-cell interference and maximize Quality of Experience (QoE).

While Deep Reinforcement Learning (DRL) methods like Proximal Policy Optimization (PPO) have shown promise, they face significant bottlenecks in this domain:

Poor Sample Efficiency and Generalization: Conventional online RL struggles with the combinatorial complexity of multi-cell selection.
Scalability Issues: Changes in network topology (e.g., increasing the number of Base Stations (BSs) or UEs) alter the state and action spaces, necessitating costly retraining of the policy.
Limited Adaptability: Existing solutions often require separate agents for different network configurations or rely on iterative environment interactions that are slow to converge in highly dynamic settings.

Methodology: Prompt Decision Transformer (PromptDT)
The authors propose a Prompt Decision Transformer (PromptDT) framework, which reformulates the multi-cell selection problem as a sequence modeling task rather than a traditional value-function or policy optimization problem.

Offline Learning & Sequence Modeling: Unlike online RL, PromptDT learns decision policies directly from offline historical trajectories. It utilizes a Transformer architecture (specifically a decoder-only GPT-style model) to perform autoregressive next-token prediction. The input sequence consists of Return-to-Go (RTG), states, and actions.
Multi-Task Learning via Prompts: To handle diverse network configurations (varying numbers of UEs, BSs, and scheduler policies like Resource-Fair (ReF) and Proportional-Fair (PF)), the framework employs task-specific trajectory prompts.
- During training, the model is exposed to a mix of tasks. A short "prompt" trajectory ( $K^*$ ) representing a specific task configuration is concatenated with a longer historical trajectory ( $K$ ).
- The model learns a unified policy representation that can generalize across these variations.
Handling Variable Dimensions: To accommodate different state/action space dimensions across tasks (e.g., 3 BSs vs. 7 BSs), the authors apply zero-padding to state and action representations to match maximum dimensions. Action masking is used during training and inference to ensure the model only selects valid actions (e.g., preventing connections to BSs with SNR below the threshold).
Few-Shot Adaptation: During evaluation, the model selects a high-reward prompt trajectory from a demonstration set for a specific unseen task. It then generates actions in a closed-loop manner conditioned on this prompt and the interaction history, allowing for adaptation to new configurations without retraining or fine-tuning.

Key Contributions

Multi-Task PromptDT Framework: The paper introduces a framework for complex multi-cell selection in CoMP where tasks vary in UE/BS counts and scheduler policies. It extends previous work to multi-discrete action spaces using masked cross-entropy loss and action masking.
Robust Few-Shot Generalization: The authors demonstrate that the framework can adapt to unseen network configurations using task-specific trajectory prompts, achieving effective generalization without the need for retraining.
Scalability and Performance: Extensive experiments show that PromptDT scales effectively with increasing task diversity and model capacity, outperforming baselines significantly.

Experimental Results
The framework was evaluated using the open-source mobile-env simulation platform across 12 distinct tasks involving varying scales (Tiny to Large) and scheduler policies.

Performance vs. PPO: PromptDT consistently outperformed single-task PPO baselines, particularly as model capacity increased (tested with 5M, 15M, and 45M parameters).
- In small-scale environments, gains were modest (5–10%).
- In larger, more complex environments (Medium to Large scales), PromptDT improved rewards by 12–30%.
- Overall, the framework achieved up to a 49% improvement in QoE in multi-task settings compared to baselines.
Unseen Task Adaptation: When evaluated on tasks not seen during training, PromptDT with higher model capacities approached or exceeded the performance of PPO agents that were specifically trained on those target environments.
Efficiency: The models demonstrated practical feasibility for real-time deployment, with inference latencies of approximately 11 ms on low-end GPUs and memory requirements of up to 90 MB (FP16).

Significance and Claims
The paper claims that PromptDT offers a scalable, unified solution for RRM in dynamic wireless networks. By leveraging the sequential modeling strengths of Transformers and the adaptability of prompt-based learning, the framework overcomes the inherent limitations of conventional RL, such as training instability and the prohibitive costs of retraining for every new network configuration.

The authors position this work as a step toward AI-native applications in wireless networks, where a single unified model can generalize across multiple operational areas and task configurations, replacing the need for maintaining separate agents for each scenario. The results suggest that increasing model capacity directly correlates with improved robustness and generalization in high-dimensional, multi-task wireless environments. Future work is suggested to explore techniques like Mixture-of-Experts to further mitigate gradient conflicts and enhance task scalability.