Researchers at Universidad Politécnica de Madrid and Universidad Autónoma de Madrid are applying reinforcement learning to design quantum repeater protocols for networks more complex than simple linear chains. The team’s work addresses a fundamental challenge in scaling quantum networks, where the number of possible protocols expands exponentially with network size. Their approach optimizes protocols beyond those relying on entanglement routing, finding solutions that outperform the widely used swap-ASAP protocol. These results indicate that routing methods should be replaced with non-routing methods, and suggest a path toward designing a sensible Quantum Internet architecture.
Reinforcement Learning Optimizes Non-Chain Quantum Repeater Protocols
Researchers are now using artificial intelligence to optimize quantum repeater protocols for network topologies beyond simple linear chains. This approach directly addresses limitations inherent in existing entanglement routing methods, which often rely on simplifying assumptions to manage computational complexity. The team’s work challenges the prevailing assumption that routing protocols represent the optimal strategy for quantum networks. They demonstrate that protocols developed using reinforcement learning outperform a routed extension of the swap-ASAP protocol, a widely recognized standard, even within parameter ranges where swap-ASAP typically excels.
This finding suggests that restricting optimization to routing protocols can inadvertently exclude superior solutions from consideration, and that a broader search of the protocol space is necessary. Deep reinforcement learning, as employed in this study, models a quantum repeater policy as a neural network that maps network states to actions.
This agent interacts with a simulated environment representing the quantum network, learning to optimize protocol parameters through trial and error. The simulation incorporates realistic imperfections, including decoherence in quantum memories and inhomogeneity in operational parameters, to ensure the resulting protocols are robust and practical. The researchers focused on networks with intersecting paths, a departure from the traditional “chain scenario” that has dominated much of the early work in quantum repeaters. This optimization process explores a high-dimensional space of non-routing protocols, seeking configurations that maximize entanglement distribution efficiency.
Entanglement Swapping and Initial Quantum Repeater Designs
Entanglement swapping, a cornerstone of quantum repeaters, enables the creation of long-distance entanglement through local measurements performed within repeaters themselves. This process relies on generating short-range entanglement and then extending it via classical communication, a technique demonstrated in recent experiments performing tasks like teleportation and secure key distribution. The efficiency of entanglement swapping is now being enhanced through the application of artificial intelligence, specifically reinforcement learning, to optimize protocols for increasingly complex network topologies.
Network Model: Quantum Memories as Graph Nodes
This granular approach allows for a more detailed representation of the network’s physical infrastructure and logical state, introducing distinct edge types to define connections between qubits both within and between stations, and to indicate established entanglement. The framework defines a graph, G = (M,E), where M represents the set of memories and E the set of edges, categorized as fiber, station, or Bell edges to reflect connection type and entanglement status.
This network model accounts for realistic imperfections in entanglement generation and swapping, processes important for extending quantum communication range. The addition of a Bell edge, indicating entanglement, between qubits is not guaranteed; failure during operation results in no edge being added, reflecting the probabilistic nature of quantum operations. Visual representations of this entanglement generation process, of the study, illustrate the dynamics between memory pairs.
Entanglement Routing as a Heuristic Protocol Limitation
The efficiency of quantum repeater protocols diminishes as network complexity increases, prompting a shift away from established routing heuristics. Researchers found that optimizing for entanglement rates across network parameters presents a challenging optimization landscape, precluding the use of methods like linear programming. Instead, machine learning approaches capable of efficiently computing gradients offer a viable alternative, expanding the search beyond protocols constrained by common heuristics like entanglement routing.
This work demonstrates that protocols designed without pre-defined routes can outperform those relying on entanglement routing, even in networks where the latter is already considered high-performing. A key challenge lies in the exponential growth of possible protocols for creating end-to-end entanglement as network size increases; previous work has often relied on heuristics to manage this complexity.
Entanglement routing simplifies protocol design by dividing the network into disjoint paths connecting target nodes, but this approach can inadvertently hinder progress. The study states, “If entanglement routing is used in a network with many intersecting routes, the resulting protocol may be outperformed by non-routing strategies, such as Reinforce.”
Reinforcement Learning Addresses High-Dimensional Protocol Spaces
Current approaches to quantum communication increasingly use machine learning to navigate the complexities of scaling networks beyond simplified models. Researchers are now applying reinforcement learning to optimize protocols in non-linear networks, expanding beyond the limitations of entanglement routing and exploring a significantly larger protocol space. This work demonstrates protocols that surpass the performance of routed extensions of the swap-ASAP protocol, even in scenarios where the latter is considered a leading standard. Machine learning, specifically deep reinforcement learning, offers a potential solution by effectively addressing the “curse of dimensionality” inherent in these complex systems.
Non-Routing Policies Outperform Routed Swap-ASAP Protocols
This performance difference establishes a basis for discarding routing strategies in favor of non-routing methods for quantum communication. The research team optimized protocols within a larger, non-routing space, moving beyond the limitations of approaches based on linear network models. The study’s findings demonstrate superior performance across networks featuring multiple intersecting paths, a characteristic that highlights the suboptimality of entanglement routing in complex topologies. The results consistently indicated that the non-routing protocols generated through reinforcement learning outperformed the routed swap-ASAP protocol, reinforcing the claim that optimal entanglement distribution does not require pre-determined pathways.
Logarithmic Negativity Benchmarks Entanglement Rate Performance
This difference in entanglement rate stems from the agent’s ability to establish end-to-end entangled pairs even when traditional routing methods fail. The study highlights that in networks with numerous intersecting routes, reliance on entanglement routing can lead to suboptimal performance. Researchers demonstrated this by correlating logarithmic negativity rates with a proposed accumulated non-commuting score for each network. The network model used in the study incorporates three key processes central to entanglement distribution in telecommunication networks: entanglement generation, swapping and decoherence.
All these processes were allowed to be faulty and inhomogeneous throughout the network, reflecting realistic conditions. The team selected the rate of logarithmic negativity as a key performance metric, using it to evaluate the effectiveness of the learned protocols. This figure summarizes a central finding: that non-routing strategies can offer significant advantages in complex network topologies.
Complex Network Parameters Modeled for Realistic Simulations
Network performance varied significantly under simulated conditions designed to mimic real-world quantum communication challenges, with the machine learning agent consistently exceeding the performance of a standard swap-ASAP protocol. The simulations incorporated inhomogeneous network parameters, specifically varying probabilities for heralding generation and swapping between 0. The complete layouts, parameters, and agent weights are available in a public repository to facilitate reproducibility and further research.
The team observed that the agent trained on a specific network configuration achieved a score around 3, while the five-station chain yielded zero advantage, as expected given the equivalence of performance between the agent and routing in that scenario. The researchers state, “The network with zero score is a five station chain and, as expected, the trained agent has the same performance as the routing strategy, resulting in zero advantage.”
👉 More information
🗞 Reinforcement Learning designs Quantum Repeater Protocols on Networks beyond the Chain Scenario
✍️ Andrés Agustí Casado, Álvaro Troyano Olivas, Javier Faba García and Laura Ortiz (Center for Computational Simulation, Universidad Politécnica de Madrid); Luis M. Robledo (Department of Theoretical Physics, Universidad Autónoma de Madrid); Vicente Martín
🧠 ArXiv: https://arxiv.org/html/2611.21414v1




See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
