Evolutionary Reservoir Computing

• Evolutionary Reservoir Computing is a framework that uses evolutionary algorithms to optimize reservoir architectures, shifting focus from weight training to design.
• Methodologies range from grid searches to genetic algorithms optimizing hyperparameters, topology, and physical substrate configurations.
• Applications include temporal prediction, speech recognition, and meta-learning, with evolved reservoirs demonstrating improved memory, robustness, and transferability.

Evolutionary reservoir computing is the branch of reservoir computing in which the reservoir is treated not as a purely random fixed recurrent substrate, but as an object of evolutionary, bio-inspired, or other black-box optimization over topology, dynamics, hyperparameters, physical operating point, or input-output interface, while the basic reservoir-computing division between a nonlinear dynamical core and an easily trained readout is usually retained (Basterrech et al., 2022). In this formulation, the central problem is shifted from recurrent weight training to reservoir design. That shift is especially consequential for physical and unconventional reservoirs, where internal parameters may be difficult to tune analytically, not fully known, or inaccessible to gradient-based optimization (Vrugt, 2024).

1. Reservoir-computing foundations and the ERC design problem

Reservoir computing begins from a recurrent nonlinear dynamical system driven by external inputs, with only a readout layer trained. In a standard ESN-style formulation, the reservoir evolves as

$x(t)=f\big(Wu(t)+\mathbf{W}^{\rm r}x(t-1)\big),$

and the output is computed as

$y(t)=g(\mathbf{W}^{\rm o}x(t)),$

or, in a common augmented form,

$y(t)=g(W[u(t);x(t)]).$

The corresponding difficulty is no longer recurrent credit assignment, but choosing a reservoir whose dynamics provide sufficient separation, memory, and stability for the task (Basterrech et al., 2022).

That design problem is the entry point for evolutionary methods. The survey literature emphasizes that RC performance depends strongly on reservoir size, spectral radius, sparsity or connectivity density, injected noise level, activation-function parameters, leak rate, feedback, and topology, and that these are still often chosen by experience, grid search, or brute force rather than principled optimization (Basterrech et al., 2022). In physical RC, the motivation is even stronger: the reservoir may be a device or material whose governing parameters are hard to differentiate through, expensive to model, or only partially observable, so derivative-free search becomes natural (Vrugt, 2024).

From an evolutionary perspective, the reservoir can therefore be optimized at several levels. One may optimize global ESN hyperparameters, recurrent topology, subsets of recurrent weights, modular or hierarchical organization, input injection schemes, delay and multiplexing parameters, observation schemes, or the operating regime of a physical substrate (Basterrech et al., 2022). A broader biological interpretation goes further: RC may emerge where rich recurrent dynamics are already available and only simple readouts need adaptation, although the long-term evolutionary stability of such generic reservoirs is itself task- and cost-dependent (Seoane, 2018).

2. Search spaces, objectives, and representations

The search space in evolutionary reservoir computing is heterogeneous. It contains discrete variables such as topology and connectivity pattern, real-valued variables such as spectral radius or leak rate, and substrate-specific control variables such as delay, input masking, chemical inflow, or sensor placement (Basterrech et al., 2022). For physical reservoirs, the relevant variables may include recurrent coupling strengths or their distributions, input injection schemes, nonlinear node characteristics, time constants, delay parameters, observation variables, operating points near or away from instabilities, multiplexing parameters, and preprocessing and sampling rates (Vrugt, 2024).

Objective functions vary accordingly. The dominant pattern in the surveyed literature is still direct optimization of predictive or classification error, typically quadratic error, although multi-objective optimization over predictive accuracy, memory capacity, robustness, complexity, and speed is repeatedly identified as an open direction (Basterrech et al., 2022). More structural criteria also appear. In the evolutionary interpretation of RC as a trade-off between separability and generalization, separability is measured by

$r^S \equiv rank\{X^{S}\},$

while generalization is measured by

$r^G \equiv rank\{X^{G}\},$

with good reservoirs requiring high $r^S$ but low $r^G$; this is explicitly presented as a conflicting objective pair and linked to Pareto reasoning (Seoane, 2018).

Several ERC systems optimize proxies for reservoir geometry rather than end-task loss alone. In a hydrodynamic reservoir, MAP-Elites optimizes the determinant of a square readout matrix,

$F(\mathcal{H}; \mathcal{D}) = | \det R |,$

as a finite-sample separability criterion, on the premise that a nonsingular $R$ supports linear decodability of the training observations (Pierro et al., 2023). In an evolved critical NCA reservoir, the objective is not task error but a criticality score built from avalanche-distribution fits to power laws, with fitness assembled from transformed $R^2$, KS, bin-occupancy, and log-likelihood terms (Pontes-Filho et al., 4 Aug 2025). In a KS-forecasting study, the genetic algorithm minimizes a composite prediction score

$y(t)=g(\mathbf{W}^{\rm o}x(t)),$0

which rewards both low error and broad spatial coverage of acceptable prediction quality (Dehghani, 22 Jun 2026).

Representation matters as much as objective choice. Some systems encode candidate reservoirs directly as hyperparameter vectors; others use graph genomes, lookup tables, or compressed generative descriptions. EARLY, for example, represents multi-reservoir ESNs as graph-based genomes containing an adjacency matrix and per-reservoir hyperparameter rows (Testu et al., 19 May 2026). EvoESN instead fixes the sparse recurrent support and evolves the active reservoir weights indirectly in a DCT basis, using low-frequency Fourier coefficients as the genotype and reconstructing the recurrent matrix by inverse DCT (Basterrech et al., 2022). This suggests that ERC is not one optimization problem but a family of coupled representation-and-fitness design problems.

3. Software ERC: from ESN tuning to structural selection

A major line of work treats ERC as optimization of software reservoirs, especially ESNs. EvoESN is a direct example: the reservoir topology is fixed, but the nonzero recurrent weights are encoded in Fourier space and evolved by a genetic algorithm. The method searches a much smaller space than direct recurrent-weight evolution and reports strong performance on Mackey–Glass, Lorenz, and monthly sunspot prediction, while preserving standard readout training by regularized linear regression (Basterrech et al., 2022).

A more architectural approach appears in EARLY, which evolves both topology and hyperparameters of multi-reservoir ESNs through graph-based genomes inspired by modular organization. Under a budget of 50,100 evaluated architectures, EARLY outperformed random search on all reported CogScale temporal tasks, reducing the mean task error from $y(t)=g(\mathbf{W}^{\rm o}x(t)),$1 to $y(t)=g(\mathbf{W}^{\rm o}x(t)),$2. The same study also examined transfer: architectures evolved by EARLY produced a lower mean validity error on an unseen cross-situational learning dataset, $y(t)=g(\mathbf{W}^{\rm o}x(t)),$3 versus $y(t)=g(\mathbf{W}^{\rm o}x(t)),$4 for random-search architectures (Testu et al., 19 May 2026). This suggests that modular reservoir organizations can be selected not only for source-task performance but also for broader reuse.

Evolutionary selection has also been used to study what kinds of reservoirs prediction actually favors. In a KS spatiotemporal-chaos setting, a genetic algorithm jointly evolved reservoir size, connectivity degree, spectral radius, input scaling, and readout regularization. The population-level fitness distribution shifted by about one order of magnitude, from $y(t)=g(\mathbf{W}^{\rm o}x(t)),$5 to $y(t)=g(\mathbf{W}^{\rm o}x(t)),$6, and the evolved reservoirs organized along a diminishing-return size–efficiency frontier. Structural analysis showed a conserved SBM-like spectral envelope, systematic refinement of low-eigenvalue modes, an intermediate modularity band with normalized modularity $y(t)=g(\mathbf{W}^{\rm o}x(t)),$7 and $y(t)=g(\mathbf{W}^{\rm o}x(t)),$8, and exponential decay of connection cost over generations (Dehghani, 22 Jun 2026). This is significant because the optimization did not merely find better hyperparameters; it exposed task-specific structural constraints on the recurrent substrate.

A closely related line uses biological priors. Connectome-based ESNs optimized by PSO, DE, GWO, and WOA across six species showed consistent gains over unoptimized biological baselines on memory capacity, Lorenz, NARMA-10, and Mackey–Glass. WOA produced the largest gains, including a $y(t)=g(\mathbf{W}^{\rm o}x(t)),$9 memory-capacity increase in C. elegans, from $y(t)=g(W[u(t);x(t)]).$0 to $y(t)=g(W[u(t);x(t)]).$1, and up to $y(t)=g(W[u(t);x(t)]).$2 NRMSE reduction on Mackey–Glass in the human connectome case. Random initialization on the same topology reliably underperformed biologically initialized optimization, indicating that empirical biological weights provided a meaningful inductive bias beyond topology alone (Guragain et al., 5 Jun 2026).

ERC has also moved into meta-learning. In meta reinforcement learning, reservoirs generated from hyperparameters $y(t)=g(W[u(t);x(t)]).$3 were evolved with CMA-ES, then used as fixed recurrent substrates while PPO trained only the policy. This improved partially observable control tasks, supported some locomotion tasks, and yielded partial transfer to unseen environments, especially in HalfCheetah and Swimmer (Léger et al., 2023). Here ERC functions as an outer loop over neural structure while lifetime learning remains in the inner loop.

4. Self-organizing and self-evolutionary reservoirs

Not all ERC is population-based outer-loop optimization. A distinct tradition treats the reservoir itself as structurally adaptive during operation. In the structural autonomous development reservoir computer, or sad-RC, the reservoir is an adaptive network of phase oscillators governed by

$y(t)=g(W[u(t);x(t)]).$4

with coupling evolution

$y(t)=g(W[u(t);x(t)]).$5

The development stage replaces the ordinary washout phase: inputs drive both oscillator states and couplings before readout training begins. The reported effect is a broad low-error region in the $y(t)=g(W[u(t);x(t)]).$6 plane, alignment between good predictive performance and high memory capacity, and near-complete convergence of different initial reservoirs to an almost identical developed structure under fixed task conditions (Zuo et al., 2023). In this usage, “evolutionary” refers to autonomous structural evolution in time rather than genetic search over generations.

A related study introduced an evolutionary reservoir computer in which recurrent weights and neuronal decay constants evolve under a genetic algorithm while readout weights are trained by ridge regression. The benchmark required separation of mixed spatial and temporal input information. After evolution, the system reached about $y(t)=g(W[u(t);x(t)]).$7 accuracy on both spatial and temporal outputs, and neurons in the internal output layer became functionally differentiated: the correlation between neuron-wise spatial and temporal mutual information fell to about $y(t)=g(W[u(t);x(t)]).$8, whereas the input-layer correlation remained weak. Evolution also reshaped the reservoir toward stronger feedforward transmission from the internal input layer to the internal output layer and weaker feedback (Yamaguti et al., 2020). This demonstrates that ERC can be used not only to improve performance but to induce specialized internal coding.

These adaptive models broaden the meaning of ERC. In one interpretation, ERC denotes evolutionary or bio-inspired outer-loop search over fixed reservoirs. In another, it denotes reservoirs that develop or reorganize their own structure online through local rules. The latter is especially relevant where task statistics unfold over time or where reservoir organization itself is part of the computational mechanism (Zuo et al., 2023).

5. Hardware, physical, and in-materio evolutionary RC

Physical RC is one of the strongest drivers of ERC, because physical substrates are often black boxes whose internal parameters are difficult to differentiate through. A hardware-oriented example is a single-node delay reservoir implemented on FPGA, where a genetic algorithm optimized the dynamical parameters $y(t)=g(W[u(t);x(t)]).$9 and PCA-based mask adaptation reduced input dimensionality. Under a population size of 20 and 40 generations, the GA required 800 evaluations; a comparable four-dimensional exhaustive search would take $r^S \equiv rank\{X^{S}\},$0 times longer. On spoken-digit recognition, GA optimization produced $r^S \equiv rank\{X^{S}\},$1 without PCA-enhanced masking and $r^S \equiv rank\{X^{S}\},$2 with it (Penkovsky et al., 2018).

In an asynchronous FPGA substrate of recurrent Boolean LUTs, evolution acted directly on the node functions rather than only on a readout. The Boolean functions implemented by CFGLUT5 units were evolved on hardware, while topology remained fixed. On a compressed MNIST classification task this yielded an accuracy improvement of about $r^S \equiv rank\{X^{S}\},$3, and the system processed inputs at over three million samples per second. The same platform also demonstrated evolvable memory and dynamic output generation (Norman-Tenazas et al., 2024). This is a particularly direct form of hardware ERC because the evolved object is the physical recurrent substrate itself.

Evolution-in-materio appears explicitly in a hydrodynamic reservoir modeled by the KdV equation. There the internal fluid dynamics were fixed, but MAP-Elites evolved the interface: cnoidal-wave encoding parameters and readout times. On XNOR, the best determinant-based separability score improved from $r^S \equiv rank\{X^{S}\},$4 to $r^S \equiv rank\{X^{S}\},$5, about three orders of magnitude, and on sigmoid regression the determinant-based fitness correlated strongly with lower out-of-sample MSE, with $r^S \equiv rank\{X^{S}\},$6, $r^S \equiv rank\{X^{S}\},$7, and $r^S \equiv rank\{X^{S}\},$8 (Pierro et al., 2023). This suggests that in physical ERC, evolving how a substrate is driven and observed can be as important as evolving its internal physics.

Biomolecular ERC extends this logic to DNA-coded substrates. A simulated DNA-bead reservoir encoded network connectivity in DNA-like strings and optimized topology by genetic algorithm. Reservoir performance depended strongly on topology; among 1000 random Erdős–Rényi graphs only $r^S \equiv rank\{X^{S}\},$9 lay in the top $r^G \equiv rank\{X^{G}\},$0 for all three benchmark tasks simultaneously. Directed evolution improved task-specific performance relative to random selection, and short sequential task evolution increased the fraction of reservoirs that performed well on multiple tasks while retaining performance on prior tasks (Pandey et al., 4 Sep 2025). A complementary contribution is infrastructural: ChemReservoir introduced a general open-source framework for chemically inspired reservoirs, with a two-level GA over cycle-based topologies and network parameters, and reported stable performance across configurations in memory-capacity tasks (Yirik et al., 31 May 2025).

Several substrate papers are not themselves ERC implementations but are important because they expose compact, evolvable design spaces. Cellular automaton reservoirs replace recurrent neural networks with simple local binary dynamics; the design variables include rule choice, neighborhood size, random mappings, CA size, and iteration depth, which the paper explicitly describes as a space that can be searched or evolved more easily than dense recurrent weight matrices (Yilmaz, 2014). A molecular reservoir based on three coupled deoxyribozyme oscillators likewise did not use evolutionary algorithms, but it identified a physically meaningful space of oscillator count, coupling topology, inflow, efflux, and observable species that later work could optimize or evolve (Goudarzi et al., 2013).

6. Criticality, biological interpretation, and open problems

Criticality is one of the most persistent but also most contested themes in ERC. Many reservoir studies associate good computation with operation near instability or the edge of chaos, because ordered regimes forget too quickly and chaotic regimes degrade reliable separation. Yet the contemporary survey literature immediately qualifies that heuristic: operating near instability can also worsen performance, and Jaeger’s criticism that the “edge of chaos” or criticality requirement is a “myth” is explicitly noted as neither mathematically well-defined nor empirically universal (Vrugt, 2024). This suggests that ERC should optimize directly for task performance and robustness rather than hard-code criticality as a universal objective.

Even with that caveat, criticality-oriented ERC has produced concrete results. An evolved critical NCA reservoir optimized by CMA-ES for avalanche power laws achieved a perfect score on the 5-bit memory task and reached MNIST accuracy competitive with the best ECA reservoirs: a maximum of $r^G \equiv rank\{X^{G}\},$1 and a mean of $r^G \equiv rank\{X^{G}\},$2, compared with $r^G \equiv rank\{X^{G}\},$3 maximum and $r^G \equiv rank\{X^{G}\},$4 mean for ECA rule 30 (Pontes-Filho et al., 4 Aug 2025). In CA-based RC more broadly, class 3 rules with random-looking behavior were reported as most effective for memory tasks, and rules 90 and 150 were singled out as especially attractive because their additive structure enables symbolic Boolean composition in reservoir space (Yilmaz, 2014). In sad-RC, autonomous development was interpreted as broadening the edge-of-chaos from a line into a wider critical domain organized around quasi-synchrony (Zuo et al., 2023). These results do not settle the controversy, but they show that criticality remains a productive search principle.

The biological interpretation of ERC adds a different set of constraints. A conceptual review proposes an evolutionary morphospace with axes of dynamical cost $r^G \equiv rank\{X^{G}\},$5, task lifetime $r^G \equiv rank\{X^{G}\},$6, and fitness-landscape ruggedness $r^G \equiv rank\{X^{G}\},$7, arguing that RC should be favored where $r^G \equiv rank\{X^{G}\},$8 is low, $r^G \equiv rank\{X^{G}\},$9 is low, and $r^S$0 is high (Seoane, 2018). On this view, reservoirs may emerge easily because many biological systems already have rich nonlinear recurrent dynamics, but full RC may be evolutionarily unstable: when tasks become few and persistent, or signaling costs rise, selection should favor specialized hard-wired circuits over metabolically costly generic reservoirs (Seoane, 2018). This perspective is highly relevant to ERC because it frames reservoir design as a problem of adaptive trade-offs rather than raw benchmark performance alone.

Open problems remain extensive. Survey work highlights the need to go beyond global hyperparameters toward projection dynamics and Lyapunov-oriented criteria, to move from single-objective error minimization toward multi-objective optimization over accuracy, memory, robustness, complexity, and speed, and to explore deep, hierarchical, interconnected, and lifelong-learning reservoir systems (Basterrech et al., 2022). Physical RC adds substrate-specific questions about black-box optimization, robustness to variability and noise, and efficient evaluation under hardware constraints (Vrugt, 2024). Recent modular ESN work also raises a more ambitious possibility: some tasks may induce architectures that are not only effective locally but reusable across broader temporal problem classes (Testu et al., 19 May 2026).

Taken together, these strands define evolutionary reservoir computing less as a single algorithmic family than as a design paradigm. Its common premise is that the reservoir is not sacrosanct. It can be searched, adapted, evolved, regularized, modularized, interfaced, or physically directed toward regimes that improve memory, separation, robustness, or transfer. The resulting field spans ESN hyperparameter evolution, compressed recurrent-weight encodings, connectome-informed search, adaptive self-developing reservoirs, hardware-aware genetic optimization, and in-materio evolution of physical substrates.