Reinforcement learning is a family of methods that learn a decision policy by trial and reward rather than from labelled examples, and across the 2026 literature it is being used less as an AI research topic in itself and more as a general-purpose replacement for hand-written heuristics in wireless precoding, quantum compilation, materials search and hardware parameter tuning.
Summary
Reinforcement learning (RL) trains a policy: a function that maps an observed state to an action, optimised to maximise a cumulative reward signal. Unlike supervised learning there is no ground-truth label for each input; the learner must explore, observe consequences, and attribute credit backwards through a sequence of decisions. The two dominant families are value-based methods, which learn the expected return of state-action pairs and act greedily (Deep Q-Networks), and policy-gradient methods, which parameterise the policy directly and adjust it along the gradient of expected reward (Proximal Policy Optimization, twin delayed deep deterministic policy gradient, normalised advantage functions). Both appear repeatedly in the sources, sometimes compared head to head within one problem.
What makes RL attractive to engineers is structural, not cognitive. Many design problems are non-convex, sequential and state-dependent, so classical optimisation either does not apply or must be re-solved from scratch for every instance. RL instead amortises the cost: you pay once to train a policy in a simulator, then get near-instant decisions that generalise across instances. That framing recurs almost verbatim across unrelated fields. Cell-free massive MIMO precoding in O-RAN is formulated as a nonconvex throughput-maximisation problem and handed to multi-agent deep RL because centralised solutions do not scale and fully distributed ones cause interference. Continuous power allocation in an optical wireless NOMA system is described as “computationally prohibitive” to optimise dynamically, so a deep RL agent learns it instead. Quantum circuit compilation is cast as combinatorial optimisation and solved by a trained policy rather than a heuristic pass.
The parameters that decide whether it works are consistent across these papers. First, simulator fidelity: the policy is only as good as the environment it was trained in, and one source explicitly studies behaviour under domain shift. Second, action-space design, which one paper elevates to the central contribution, showing that reformulating the action space plus action masking materially changes achievable performance on the same underlying task. Third, sample efficiency and reward shaping: a quantum error-correction paper uses curriculum learning plus an analytical master-equation solution specifically to accelerate training, and a molecular agent replaces per-step energy rewards with energy- and validity-based rewards over multiple compositions to get generalisation. Fourth, inference-time cost, which becomes a hard constraint when the policy sits inside a real control loop.
A separate thread worth flagging: a 2025 theory paper argues that transformer attention is the exact solution to a degenerate one-sided entropic optimal transport problem, and that its backward pass yields gradients mathematically identical to advantage-based policy gradients from RL. The same source cautions that the claim is narrower than a general attention-optimal-transport equivalence and has been mis-cited elsewhere. If it holds, it is a statement about why the two learning paradigms share machinery, not an engineering result.
Evidence base
- 14 May 2026: an RL policy for logical-to-physical qubit mapping, with local-search post-processing, reports 65-85% reduction in SWAP overhead versus existing quantum compilers on MQTBench and Queko circuits.
- 2 June 2026: a PPO-based QKD parameter tuner with TCN forecasting reports a 14-25% increase in median secure key rate and median QBER falling from 3.0% to 1.5%, with an exploratory quantum RL extension claiming 29.2% throughput gain on E91.
- 15 May 2026: entangled quantum multi-agent RL reaches 0.854 win rate on CHSH against a proven classical ceiling of 0.75, and about 0.85 versus 0.40 success on cooperative navigation against classical MAA2C, while the same paper criticises the field for evaluating without provable baselines.
- 28 May 2026: an agent trained exclusively online with no pretraining discovers up to an order of magnitude more valid isomers on unseen chemical formulas than single-composition RL baselines using per-step energy rewards.
- 5 May 2026: on a physical superconducting QPU, a single-qubit hybrid agent solves CartPole in substantially fewer episodes than a comparable classical actor-critic, but stability depends strongly on both inference control frequency and measurement shot count, and a real-time control demonstration remains future work.
- 11 August 2025 (revised 29 December 2025): a single-author paper proves attention’s forward pass is the exact solution to a degenerate one-sided entropic optimal transport problem and argues its backward pass yields gradients identical to advantage-based policy gradients, while warning the claim is narrower than commonly cited.
Open questions
- Do the reported margins survive independent reimplementation against strong classical baselines rather than the random or heuristic incumbents used in the source papers, particularly the 65-85% SWAP reduction?
- How much do these policies degrade under domain shift between the training simulator and measured hardware or channel conditions, which the emitter-localisation work raises but does not settle?
- Can learned controllers meet hard real-time and formal-constraint requirements, given that inference frequency and shot budget already dominate stability in the only physical-hardware study here?
- How much of the performance is attributable to RL rather than to problem-specific engineering of the action space, masking and reward, which one paper treats as the primary lever?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.