Laser Chaos-Based Multi-Processing Learning

• LC-ML is a photonic learning framework that harnesses chaotic laser dynamics to process multiple decisions concurrently across various dimensions.
• It utilizes physical substrates like semiconductor lasers and multimode fibers to achieve rapid adaptation, synchronization, and effective exploration in reinforcement tasks.
• The architecture demonstrates scalable multi-processing with threshold-based adaptations and coordinated decision-making, enabling applications from underwater relays to high-dimensional bandit problems.

Laser Chaos-Based Multi-Processing Learning (LC-ML) denotes a class of photonic learning systems in which chaotic laser dynamics provide the physical substrate for exploration, nonlinear transformation, synchronization-based coordination, or stochastic decision formation, while multiple decisions, agents, channels, or tasks are processed concurrently in time, wavelength, space, or network topology. The label is explicit in underwater acoustic relay assignment, and a broader LC-ML family is suggested by related work on laser-chaos bandits, parallel bandit architectures for reinforcement learning, multimode mode-competition learning, synchronization-mediated joint decisions, and programmable chaos-on-comb platforms (Chen et al., 8 Jul 2025, Naruse et al., 2018, Urushibara et al., 2022, Ito et al., 2023, Shi et al., 4 Jan 2026).

1. Historical emergence and conceptual scope

The modern LC-ML lineage in arXiv-visible form begins with ultrafast photonic reinforcement learning for the two-armed bandit, where a semiconductor laser with delayed optical feedback was sampled at up to 100 GSample/s and combined with a tug-of-war threshold rule to achieve a maximum adaptation speed of 1 GHz; the same study identified an exact coincidence between a negative autocorrelation minimum in the laser chaos and maximal decision-making performance (Naruse et al., 2017). The next step was scalability by time-division multiplexing of a single chaotic time series, which yielded experimental demonstrations of multi-armed bandits with up to 64 arms and a power-law scaling of cycles to $\mathrm{CDR}=0.95$ of approximately $55N^{1.16}$ (Naruse et al., 2018). Subsequent work addressed arm-order recognition rather than only best-arm identification by introducing confidence-interval-controlled exploration in the threshold tree, thereby improving the ranking of the top four arms while maintaining total reward close to reward-maximizing baselines (Narisawa et al., 2020).

A second development line replaced single scalar bandits by physically richer substrates. Chaotic itinerancy in multimode semiconductor lasers was used to solve multi-armed bandits with $M$ up to 513, with the number of plays to reach $\mathrm{CDR}=0.95$ following $N_{\mathrm{play}} \approx 318\,M^{0.70}$, thereby outperforming UCB1-tuned for large $M$ (Iwami et al., 2022). Parallel Bandit Architecture for Reinforcement Learning (PBRL) then recast multi-state RL as a bank of state-indexed bandits, showing faster adaptation than Q-learning on cart-pole balancing and a beneficial role for the autocorrelation structure of laser chaos (Urushibara et al., 2022). In parallel, multimode mode competition, injection-controlled dominant-mode ratios, and synchronization in coupled-laser networks extended the learning substrate from temporal chaos to mode space and network topology (Iwami et al., 2022, Ito et al., 2023, Iwami et al., 2023).

More recent work broadened LC-ML in two directions. One direction replaced full deterministic chaos by coarse-grained stochastic surrogates learned from laser observables, showing that low-pass stochastic differential equation models can reproduce reinforcement-learning-relevant statistics and comparable bandit performance (Fukushi et al., 6 Jun 2025). The other direction expanded the hardware substrate itself: spatiotemporal chaos in graded-index multimode fibers was used as a photonic neural network for classification (Kesgin et al., 2024), and a programmable “chaos-on-comb” architecture generated 50 spectrally sliced chaotic channels and supported a 256-armed bandit with convergence-cycle scaling $CC = 24.55\,N^{0.86}$ (Shi et al., 4 Jan 2026). This trajectory suggests that LC-ML is not a single algorithm, but a hardware-centered family of learning architectures built around chaotic photonic dynamics.

2. Physical substrates and dynamical primitives

The physical basis of LC-ML lies in the broader physics of laser diode chaos. Delayed optical feedback, optical injection, direct modulation, optoelectronic feedback, and mode competition are all established routes to chaos in semiconductor lasers, with the linewidth-enhancement factor $\alpha$, delay $\tau$, and feedback strength playing central roles in setting dimensionality, bandwidth, and sensitivity (Sciamanna et al., 2015). In this sense, LC-ML uses not merely light as an analog carrier, but specific nonlinear dynamical regimes of laser devices as trainable or biasable computational media.

One widely used primitive is leader-laggard switching in mutually coupled lasers. In the two-laser setting, short-term cross-correlations $C_1(t)$ and $55N^{1.16}$0 determine which laser leads, and the leader identity can be mapped directly to an arm choice; the switching is spontaneous, irregular, and approximately delay-scaled, making it suitable as a physical exploration process (Fukushi et al., 6 Jun 2025). A network generalization appears in the four-laser, two-player, two-slot system where lag synchronization within each player pair provides leader-based action selection and zero-lag synchronization across complementary slot pairs enforces conflict avoidance (Ito et al., 2023). In that architecture, the same delayed nonlinear network simultaneously supplies local exploration and cross-agent coordination.

A second primitive is chaotic mode competition in multimode lasers. In a multimode Fabry–Perot semiconductor laser with optical feedback and injection, the dominant mode is the longitudinal mode with maximum instantaneous intensity, and its dominant-mode ratio acts as an empirical selection probability. Optical injection strength and detuning control this dominant-mode ratio, but the control law is mode dependent because each mode experiences a different optical feedback phase (Iwami et al., 2022). Experimentally, positive wavelength detuning was found to produce efficient mode concentration to one longitudinal mode with a small optical injection power, and this fast concentration translated into a fast convergence of the correct decision rate in a four-armed bandit task (Iwami et al., 2023). An important clarification from the multimode literature is that the region of large dominant-mode ratio does not correspond to the injection-locking range; in multimode lasers with optical feedback and injection, mode dominance and injection locking are distinct dynamical phenomena (Iwami et al., 2022).

A third primitive is chaotic itinerancy. In multimode semiconductor lasers with external feedback and per-mode optical injection, the instantaneous dominant longitudinal mode hops among unstable quasi-attractors, with mode-dependent residence times that can be biased by injection. This provides a natural mapping from modes to arms in a large-scale bandit, where unforced itinerancy supplies exploration and optical injection supplies exploitation (Iwami et al., 2022). A related, but spatial rather than modal, primitive appears in graded-index multimode fibers driven by a picosecond laser and an OAM phase. There, excitation of high-order modes and strong Kerr nonlinearity generate spatiotemporal chaos, which acts as a high-dimensional photonic reservoir (Kesgin et al., 2024).

3. Learning mechanisms and computational mappings

The earliest LC-ML decision rules are threshold based. In the two-armed case, a sampled chaotic signal is compared with a threshold determined by a scalar threshold adjuster $55N^{1.16}$1; reward and failure then shift the threshold so that future comparisons bias choices toward the empirically better arm (Naruse et al., 2017). For $55N^{1.16}$2 arms, the threshold logic becomes a binary tree: bit $55N^{1.16}$3 of the arm index is decided by comparing a later sample of the same chaotic waveform to a threshold $55N^{1.16}$4, where the threshold branch depends on the preceding bits. Learning is then local to the traversed path in the tree, but statistically global because each threshold aggregates reward information over a subset of arms (Naruse et al., 2018).

Arm-order recognition extends this mechanism by attaching confidence intervals to the two subsets of arms separated by each threshold. If the confidence intervals overlap, the threshold step sizes are reduced, preserving exploration; if they do not overlap, the step sizes are increased, accelerating exploitation. The result is a hierarchy of adaptive exploration rates defined over subsets $55N^{1.16}$5 and $55N^{1.16}$6, rather than over individual arms only (Narisawa et al., 2020). This is a distinctive LC-ML pattern: the physical stochastic source remains unchanged, but the learning rule modulates how aggressively different partitions of the action space are separated.

In PBRL, the same threshold principle is lifted from stateless bandits to multi-state reinforcement learning. Each environment state $55N^{1.16}$7 is associated with its own threshold $55N^{1.16}$8, and action selection is again a threshold comparison against a chaotic laser sample. Success updates only the current state's threshold, whereas failure retrospectively updates all thresholds visited in the episode with a discounted correction. This couples many bandit-like units through trajectory-level reinforcement while preserving a simple physical comparator-based action mechanism (Urushibara et al., 2022). The architecture thereby unifies policy parameter and action-selection mechanism in a way that is physically natural for photonics and unlike tabular Q-learning.

Some LC-ML systems use direct reward feedback; others do not. In the four-laser conflict-free joint decision network, learning of arm probabilities is implicit: slot hit probabilities are static and known for analysis, the laser network does not directly receive reward signals optically or electronically, and no explicit UCB/$55N^{1.16}$9-greedy/Thompson update runs on the fly (Ito et al., 2023). By contrast, in chaos-on-comb decision making the stochastic and deterministic components are written explicitly as

$M$0

$M$1

with $M$2 extracted from chaotic optical channels and $M$3 updated by a tug-of-war rule from wins $M$4 and losses $M$5 (Shi et al., 4 Jan 2026). This formulation makes explicit a recurring LC-ML decomposition: physical chaos supplies exploration, while low-dimensional adaptive biases encode exploitation.

A further abstraction is the replacement of deterministic laser chaos by learned stochastic surrogates. Generator Extended Dynamic Mode Decomposition (gEDMD) was used to fit two-dimensional stochastic differential equation models on short-term cross-correlation observables $M$6. The resulting low-pass models preserve the peak shift in leadership bias and the power-law-plus-peak structure of switching-time statistics, and they match deterministic-chaos bandit performance closely: $M$7 for the original deterministic chaos and $M$8 for the stochastic low-pass model in a $M$9 two-armed bandit (Fukushi et al., 6 Jun 2025). This suggests that, for some LC-ML tasks, the operationally relevant object is not full ultrafast chaos itself but a coarse-grained stochastic law over decision variables.

4. Multi-processing architectures and coordination mechanisms

The “multi-processing” component of LC-ML takes several distinct architectural forms. In coupled-laser synchronization networks, it means multiple agents or players embedded in a common physical graph. In PBRL, it means many state-indexed bandits updated in parallel. In chaos-on-comb systems, it means many spectrally sliced chaotic channels processed concurrently. In underwater relay assignment, it means multiple source nodes simultaneously learning and exchanging assignments (Ito et al., 2023, Urushibara et al., 2022, Shi et al., 4 Jan 2026, Chen et al., 8 Jul 2025).

The four-laser joint-decision network is the clearest example of physically enforced coordination. Four distributed-feedback semiconductor lasers are arranged so that lasers 1A–1B and 2A–2B form two player pairs, with lag synchronization inside each pair and zero-lag clustering across complementary slot assignments. The resulting joint decision table overwhelmingly yields $\mathrm{CDR}=0.95$0 or $\mathrm{CDR}=0.95$1, rarely $\mathrm{CDR}=0.95$2 or $\mathrm{CDR}=0.95$3, so conflict avoidance is implemented in the network topology itself rather than by explicit negotiation (Ito et al., 2023). This is a strong LC-ML motif: constraints and cooperation can be embedded directly in the physics.

PBRL realizes a different notion of concurrency. In the cart-pole example, the continuous state is discretized into $\mathrm{CDR}=0.95$4 states, and each state is assigned a two-armed bandit with its own threshold. A single chaotic laser time series is then used as a common exploration source for this entire threshold table. On failure, many thresholds are updated at once according to the episode trajectory, so learning is distributed over a large set of simple local units rather than concentrated in a monolithic value function (Urushibara et al., 2022). This is not multi-agent control, but it is multi-process learning in a literal sense.

The underwater acoustic relay assignment system introduces the LC-ML term directly and couples chaos-driven exploration with a multi-processing exchange process. Each source node uses laser-chaos-generated random numbers and adaptive thresholds to form a relay preference list; then multiple requesters simultaneously attempt relay exchanges under either Classical Stable Arrangement (CSA) or Ambiguous Stable Arrangement (ASA) rules. The multi-processing stage accelerates convergence toward stable high-throughput assignments, while the ambiguity threshold $\mathrm{CDR}=0.95$5 in ASA reduces volatility by suppressing exchanges driven only by small preference differences (Chen et al., 8 Jul 2025).

At the largest demonstrated scale, chaos-on-comb transforms a delayed-feedback chaotic laser into a programmable bank of broadband channels. Spectral slicing provides up to 50 channels with preserved statistical independence, and the same platform supports a 256-armed bandit by combining parallel channels with time multiplexing (Shi et al., 4 Jan 2026). This architecture generalizes earlier time-division multiplexed bandits from a single chaos stream (Naruse et al., 2018) into a hybrid time–frequency multi-processing substrate.

5. Representative results and application domains

In cooperative decision making, the four-laser synchronization network achieved near-optimal numerical performance for the $\mathrm{CDR}=0.95$6-player, $\mathrm{CDR}=0.95$7-slot case with $\mathrm{CDR}=0.95$8, $\mathrm{CDR}=0.95$9, 100 plays per cycle, averaged over 10 cycles: player rewards were approximately 50.4 and 50.3, team reward approximately 100.7, and conflict rate approximately 0. In the fiber-coupled experiment, zero-lag synchronization between complementary lasers reached $N_{\mathrm{play}} \approx 318\,M^{0.70}$0, and the cooperative system yielded player rewards of approximately 38.0 and 57.7, team reward approximately 95.8, and conflict rate approximately 0.085; the uncoordinated baseline gave team reward approximately 79.7 and conflict rate approximately 0.60. In a $N_{\mathrm{play}} \approx 318\,M^{0.70}$1-player, $N_{\mathrm{play}} \approx 318\,M^{0.70}$2-slot numerical extension with $N_{\mathrm{play}} \approx 318\,M^{0.70}$3, $N_{\mathrm{play}} \approx 318\,M^{0.70}$4, $N_{\mathrm{play}} \approx 318\,M^{0.70}$5, the system achieved conflict-free operation and team reward approximately 100.7 over 100 plays (Ito et al., 2023).

In large-scale bandits based on chaotic itinerancy of a multimode semiconductor laser, the number of plays needed to reach $N_{\mathrm{play}} \approx 318\,M^{0.70}$6 followed $N_{\mathrm{play}} \approx 318\,M^{0.70}$7, whereas UCB1-tuned followed $N_{\mathrm{play}} \approx 318\,M^{0.70}$8. Final-play regret scaled as $N_{\mathrm{play}} \approx 318\,M^{0.70}$9 for the laser method and $M$0 for UCB1-tuned. The laser scheme was reported to be about 2.7× faster than UCB1-tuned at $M$1 and about 6.3× faster at $M$2, while operating at a 0.1 ns sampling interval (Iwami et al., 2022). Earlier ultrafast photonic RL on two arms reached a maximum adaptation speed of 1 GHz, with best performance at the sampling interval where the laser chaos exhibited maximum negative autocorrelation (Naruse et al., 2017).

In modeling and surrogate learning, the low-pass stochastic model derived by gEDMD reproduced reinforcement-learning performance of the original deterministic chaos with $M$3 versus 0.5912 for the deterministic engine, and with $M$4 versus $M$5 for the deterministic system (Fukushi et al., 6 Jun 2025). In forecasting rather than bandit learning, Locality Blended Next-Generation Reservoir Computing (LB-NGRC) trained on the Ikeda map of a chaotic laser achieved forecasting horizons exceeding five Lyapunov times and reproduced the long-term “climate” of the attractor over long rollouts (Gauthier et al., 30 Mar 2025). In a different task class, spatiotemporal chaos in graded-index multimode fibers improved classification accuracy from 75% to 83.34% on BreastMNIST, from 75.82% to 78.62% on FashionMNIST, and from 26.46% to 84.61% on EuroSAT (Kesgin et al., 2024).

The explicit LC-ML relay-assignment study in underwater acoustic networks reported that laser-chaos-based random numbers and multi-processing in the exchange process had a positive effect on higher throughput and strong adaptability with environmental changing over time, while ambiguous cognitions produced a stable configuration with less volatility compared to accurate ones (Chen et al., 8 Jul 2025). At the largest optical-chaos scale reported here, the chaos-on-comb platform produced a single-channel effective bandwidth of 543.8 GHz, a 0–500 GHz terahertz noise source with $M$6 dB excess noise ratio, and a 256-armed bandit with convergence-cycle scaling $M$7 (Shi et al., 4 Jan 2026).

6. Limitations, misconceptions, and research directions

A recurring misconception is that all LC-ML systems are fully closed-loop reinforcement learners. That is not the case. Some systems, such as the conflict-free joint-decision laser network, realize a fixed but useful stochastic policy defined by synchronization structure and intrinsic chaos, while not feeding rewards back into the photonic hardware during an experiment cycle (Ito et al., 2023). Conversely, bandit solvers based on tug-of-war or confidence-interval updates do implement explicit reward-modulated parameter adaptation (Naruse et al., 2017, Narisawa et al., 2020, Shi et al., 4 Jan 2026). LC-ML therefore spans both reward-adaptive and topology-defined physical policies.

Another misconception is that stronger mode concentration or large dominant-mode ratio is equivalent to injection locking. Multimode laser studies show explicitly that the region for large dominant-mode ratio does not correspond to the injection-locking range; high dominance can occur without locking, and locking can occur without global dominance (Iwami et al., 2022). This distinction matters for design, because locking is spectrally stabilizing while dominance is computationally decisive. Similarly, experimental mode-competition control studies emphasize that positive wavelength detuning can concentrate energy rapidly into one mode with small injection power, but this does not collapse the full bifurcation structure of the total intensity, which remains a complex mixture of periodic, quasi-periodic, and chaotic regimes (Iwami et al., 2023).

Hardware nonidealities remain central. Synchronization-based networks are sensitive to delay matching, device mismatch, and coupling-path equality; perfect conflict-free operation in simulation becomes nonzero collision rate experimentally, and player fairness degrades under parameter mismatch (Ito et al., 2023). Multimode-fiber reservoirs require careful tuning of input peak power and launch conditions because too little nonlinearity under-processes the data, while too much can produce over-processing or beam-cleaning effects (Kesgin et al., 2024). Chaos-on-comb presently relies on discrete lasers, modulators, EDFAs, long dispersive fiber, and high RF drive powers, and practical deployment will require photonic integration and faster co-designed detection electronics (Shi et al., 4 Jan 2026).

At the algorithmic level, coarse-grained stochastic surrogates reduce modeling burden but introduce their own limits: high-degree polynomial drift and diffusion terms reduce interpretability, preprocessing choices matter strongly, and naive gEDMD scales poorly in dimension (Fukushi et al., 6 Jun 2025). The underwater LC-ML relay-assignment study likewise notes a simple network model, limited evaluation metrics, and the still-open question of the optimal number of requesters as a function of network size and channel conditions (Chen et al., 8 Jul 2025). The plausible long-term direction is a shift from static, topology-defined policies toward fully learning physical networks in which chaotic dynamics, synchronization, reward feedback, and hardware programmability interact directly. That trajectory is already visible in proposals for hybrid photonic–electronic LC-ML chips, larger synchronized networks, programmable spectrum-sliced chaos sources, and stochastic surrogates that can stand in for expensive physical subsystems (Ito et al., 2023, Fukushi et al., 6 Jun 2025, Shi et al., 4 Jan 2026).