PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Behavioral Evolved Substrates Through Neuromodulation and Activation Selection

Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel

arXiv:2610.00148v1cs.NEcs.AI

TL;DR

Can evolution produce multiple behaviors from one indirectly encoded genotype, and is neuromodulation alone sufficient? The paper combines neuromodulation with task-specific activation selection. It reports 100% simultaneous five-task success, with a median of 14 generations.

  • Problem

    The paper asks whether neuromodulation alone supports simultaneous multi-task learning in indirectly encoded networks, or whether computational primitive diversity is also needed.

    Source: Introduction · pp. 1–2

  • Method

    The approach evaluates one shared network with task-specific neuromodulatory inputs and activation functions.

    Source: Network Architecture and Encoding · p. 3

  • Results

    100% 5-task success (median 14 generations) with per-task activation functions matched to parity and threshold tasks.

    Source: Conclusion · p. 9

  • Takeaways & Limitations

    The paper concludes computational primitives should be treated as evolvable traits, enabling one evolved topology to express different behaviors across tasks.

    Source: Conclusion · p. 9

    Evaluation requires explicit task identity at test time.

    Source: Limitations · p. 8

Abstract

from arXiv · show

Open-ended artificial life systems must acquire diverse competencies from a single evolving genotype. Biological brains combine neuromodulation, which reconfigures circuits without changing connections, with diverse neuron types matched to specific computational roles. Can artificial evolution achieve something analogous in indirectly encoded substrates? Using indirectly encoded substrates evolved via CPPNs, we show through more than 10,000 experiments that neuromodulation alone is insufficient: under evolutionary search, monotonic activation functions impose a 75% ceiling on parity tasks that persists regardless of capacity, topology, or population size. This is an evolutionary search barrier, not a representational limit, since Adam gradient descent achieves 100% on the identical architecture. We combine neuromodulation with per-task activation function selection, matching oscillatory primitives to parity tasks and monotonic to threshold tasks, producing multi-behavioral evolved substrates. The result: 100% simultaneous 5-task success across all 30 seeds (median 14 generations). This generalizes across the oscillatory activation class: all four functions reach 100% (30 seeds each). Neither mechanism suffices alone. The barrier extends to higher-arity and asymmetric tasks, while multi-layer depth provides an alternative path. For open-ended evolution, the computational primitive should itself be an evolvable trait. At inference, one evolved genotype expresses many behaviors.

Introduction

The paper asks whether neuromodulation alone enables indirectly encoded networks to learn multiple competencies, or whether task-matched computational primitives are also needed. It maps multi-task compatibility and investigates the evolutionary barrier and a per-task activation solution.

Introduction

  • Biological neuromodulation can reconfigure circuits without changing synaptic connections, but brains also use neuron types matched to distinct computational roles.The paper asks whether artificial systems need both mechanisms for simultaneous multi-task learning.Source: Introduction · p. 1
  • Across five Boolean tasks, the study maps how activation choice and neuromodulation interact to determine multi-task feasibility.The analysis spans more than 10,000 experiments and distinguishes parity from threshold tasks.Source: Introduction · p. 1
  • At a single layer, monotonic activations impose a 75% parity ceiling in multi-task evolutionary search, which gradient descent can bypass.Oscillatory primitives or added depth can resolve the optimizer-dependent barrier.Source: Introduction · p. 2
  • Matching activation functions to task structure enables one evolved network to achieve 100% simultaneous success across five tasks and all 30 seeds.The study also reports that the oscillatory advantage generalizes across four functions and that the barrier extends to higher-arity and asymmetric tasks.Source: Introduction · p. 2

Related Work

Prior work addresses forgetting, task interference, and activation diversity through different mechanisms. This paper distinguishes task-level changes to computational primitives from methods that modulate existing computations or diversify activations across neurons.

Related Work

  • Elastic Weight Consolidation and Synaptic Intelligence regularize weights against forgetting, but do not address activation-function mismatch.Progressive Networks avoid interference with task-specific columns, while PackNet creates task-specific subnetworks through pruning.Source: Related Work · p. 2
  • Prior neuromodulation methods target existing computations; the paper argues modulation alone cannot resolve activation-task mismatch under evolutionary search.Earlier examples include modulatory neurons, structural separation, and differentiable neuromodulated plasticity.Source: Related Work · p. 2
  • Multi-task methods address interference through gradient procedures or expert routing, whereas per-task activation changes the computational primitive used by a shared network.The proposed mechanism is analogous to mixture-of-experts routing but operates at the primitive level.Source: Related Work · p. 2
  • Earlier activation-diversity approaches assign functions across neurons; this work instead assigns one function to all neurons per task.The shared network can then switch computational primitives at inference.Source: Related Work · p. 2

Methods

The method evolves a shared network and reconfigures it for each task through neuromodulation and activation selection. The experiments use Boolean tasks and measure simultaneous convergence across multiple tasks.

Methods

  • One network shares topology and weights across tasks, while neuromodulation and per-task activation selection reconfigure its behavior.The same network is evaluated with task-specific modulatory inputs and activation functions.Source: Methods · p. 2, Network Architecture and Encoding · p. 3

Neuromodulation Mechanism

  • A four-dimensional neuromodulator vector projects through evolved receptor densities to control gain, bias, and sigmoid gating.These controls derive from the same projection and are coupled; ACh selects output identity or inversion in the experiments.Source: Neuromodulation Mechanism · pp. 2–3

Network Architecture and Encoding

  • EMR-HyperNEAT evolves a CPPN that generates the substrate topology, yielding about 10 hidden neurons for two-input Boolean tasks.Indirect encoding is used because manually specifying heterogeneous per-node activations becomes intractable as networks grow.Source: Network Architecture and Encoding · p. 3

Boolean Logic Benchmark

Activation Functions

  • The baseline uses tanh for every task, while the per-task condition assigns sin to XOR and tanh to threshold tasks.Robustness experiments extend the activation palette to five monotonic and five oscillatory functions.Source: Activation Functions · p. 4

Experimental Design

  • A run converges when every task reaches at least 98% accuracy, and convergence rates are compared across independent seeds.The study reports more than 10,000 experiments: 3,969 baseline and 6,570 robustness experiments.Source: Experimental Design · p. 4

Results

With uniform tanh, evolutionary search struggles to solve XOR alongside threshold tasks, and extending the search budget does not remove the observed ceiling. Compatibility depends on the task combination.

The Multi-Task Barrier

  • 0% of five-task runs converged under uniform tanh, while success declined from 100% for one task to 68% for two and 34% for three.Networks reached 100% on threshold tasks but plateaued at exactly 75% on XOR.Source: The Multi-Task Barrier · p. 4
  • 100% for AND+OR versus 0% for XOR+AND in the two-task benchmark, demonstrating pair-specific compatibility.Replication across 630 runs found all XOR-free conditions converged, while XOR collapsed to 0% when combined with AND or NAND.Source: The Multi-Task Barrier · p. 4
  • 1,000 generations did not break the ceiling: XOR remained at 0.750, supporting a convergence barrier rather than insufficient runtime.The replication also found XOR-free conditions converged, while XOR collapsed when paired with AND or NAND.Source: The Multi-Task Barrier · p. 4

The 75% Ceiling: Monotonicity vs. XOR

With monotonic activations, evolutionary search plateaus at 75% on XOR, while oscillatory activation matches XOR’s alternating pattern. Adam reaches 100% with uniform tanh on the same architecture, identifying the barrier as search-related rather than representational.

The 75% Ceiling: Monotonicity vs. XOR

  • Oscillatory activation matches XOR’s LOW→HIGH→LOW pattern, whereas evolution consistently fails to discover the coupled tanh weight-receptor configuration for XOR with a conflicting task.The passage distinguishes this evolutionary search failure from the ability of multiple hidden neurons to represent non-monotonic mappings.Source: The 75% Ceiling: Monotonicity vs. XOR · pp. 4–5
  • 75% is the XOR ceiling for monotonic tanh: it stays HIGH at the largest input sum, where XOR requires LOW.Oscillatory sin returns LOW there and solves XOR at 4/4 (100%).Source: The 75% Ceiling: Monotonicity vs. XOR · p. 4, Figure 2 · p. 5

The Optimizer Independence Test

  • Adam with uniform tanh reaches 100%, while evolutionary search with uniform tanh reaches 76.7%, showing an evolutionary search barrier rather than a representational limit.Adam’s gradients navigate the coupled weight-receptor space that the tested evolutionary strategy did not.Source: The Optimizer Independence Test · p. 5

Per-Task Activation Functions

Per-task activation selection pairs sin with XOR and tanh with threshold tasks while keeping weights, receptor densities, and the output neuron shared. Together with neuromodulation, this yields simultaneous success across five tasks, while ablations show that neither mechanism alone suffices.

Per-Task Activation Functions

  • Sin for XOR and tanh for threshold tasks vary by task, while weights, receptor densities, and the output neuron remain shared.Only the activation and NT vector change between task evaluations.Source: Per-Task Activation Functions · p. 5
  • 100% of seeds achieved simultaneous success on all five tasks, with median convergence in 14 generations.XOR converged last as evolution tuned sin’s oscillatory phase.Source: Per-Task Activation Functions · p. 5, Figure 3 · p. 5

Ablation Studies

  • Uniform sin also converges on this five-task Boolean benchmark, but per-task assignment is not significantly faster and is intended to match primitives to task structure.Uniform sin’s reported sufficiency is specific to this 2-input Boolean set; the supplied passage says the comparison reverses on continuous regression.Source: Ablation Studies · p. 5
  • Neither mechanism suffices alone: uniform tanh with neuromodulation and per-task activation with flat NT vectors both yield 0% success.With flat NT vectors, XOR reaches 100% but threshold tasks plateau at 50%; preserving inversion while removing gain/bias modulation also fails to converge.Source: Ablation Studies · pp. 5–6
  • A single evolved substrate reuses the same weights and receptor densities across all five tasks by switching the NT vector and activation function at inference.Source: Ablation Studies · p. 6
  • At 2-input, gain, bias, or gating alone each reaches 100%, but each converges 1.7× slower than the combined formulation.At 3-input, bias-only remains at 100% while gain-only falls to 63.3%.Source: Ablation Studies · p. 6
  • Per-task activation removes all 2-task barriers: all 10 task pairs converge, including XOR+AND, which fails under both uniform conditions.Table 2 reports 2-task convergence rates with at least 98% defined as convergence.Source: Ablation Studies · p. 6, Table 2 · p. 6

Oscillatory Class Generalization

Across the tested five-task benchmark, all four oscillatory activations reach 100%, while tanh retains the XOR ceiling. At three inputs, the oscillatory advantage persists but varies by function.

Oscillatory Class Generalization

  • All four oscillatory functions reach 100% in the five-task benchmark, while tanh reaches 76.7% and its failures show XOR stuck at 75%.The oscillatory functions converge roughly 4–6× faster than tanh in this sweep.Source: Oscillatory Class Generalization · p. 6
  • At three inputs, cos and GCU each reach 90%, dampened sin reaches 56.7%, sin(2x) reaches 40%, and tanh reaches 26.7%.The doubled frequency of sin(2x) creates more zero-crossings and a rugged fitness surface comparable to monotonic activations at higher arity.Source: Oscillatory Class Generalization · p. 6

Robustness

Robustness tests find that per-task activation remains effective across higher arities and matched comparisons, while added depth offers another route to success. The results also show boundaries across task scaling, domains, and training conditions.

Parameter sensitivity

  • Across modulation strengths from s=1.0 through s=20.0 and populations from 100 to 1,000, all conditions achieve 100% convergence.Only s=0.5 degrades, to 93.3%.Source: Robustness · p. 6

Depth and capacity

  • Two-layer uniform-tanh networks reach 100% versus 76.7% for one layer, while one-layer per-task activation is 4× faster and uses 2.9× fewer parameters.At two layers, increasing capacity from 20 to 40 neurons raises tanh success from 43.3% to 76.7%, with Parity-4 solved in all seeds.Source: Robustness · pp. 6–7

Higher-arity tasks

  • At three inputs, per-task activation reaches 100% across tested populations versus 26.7% for uniform tanh; at five inputs it reaches 83.3% versus 60% for two-layer tanh.At four inputs, per-task activation reaches 30% at Pop=1500, while two-layer tanh reaches 73.3%.Source: Robustness · pp. 6–7

Topology and asymmetric tasks

  • At three inputs, full recurrence reaches 80% versus 20% for feedforward under uniform tanh, but both topologies reach 100% with per-task activation; sin also resolves IMPLY where ReLU reaches 6.7% and ELU 0%.Source: Robustness · p. 7

Task scaling

  • Adding XNOR yields 100% at six tasks, but indirect encoding reaches 0% at seven, eight, and ten tasks; two-layer networks reach 100% at eight tasks.At ten tasks, the implication-family tasks are the sole failure points; performance ranges from 30–50% with multi-layer depth.Source: Robustness · p. 7

Frozen substrate

  • Applying sin post hoc to weights trained under tanh leaves XOR at 0% solved, but retraining from those frozen weights with per-task activation converges in two generations.The passage locates the barrier in training dynamics, not evaluation.Source: Robustness · p. 7

Domain generality

  • On smooth continuous regression, uniform tanh reaches 100% while per-task activation reaches 43.3% and uniform sin 6.7%; on Gaussian XOR, two-layer tanh reaches 93.3%.For four-class Gaussian XOR, tanh reaches 30% and per-task activation 100%; only two-layer succeeds on the 4D Iris-like synthetic task, at 80%.Source: Robustness · p. 7

Per-task activation versus uniform sin

  • Across 15 matched robustness conditions, per-task activation converges more often in 11 and less often in none; pooled success is 42.4% versus 26.9% for uniform sin.The comparison is seed-matched; per-task activation succeeds on 75 runs where uniform sin fails, versus 5 in the opposite direction.Source: Robustness · p. 7

Task Compatibility in Modulatory Space

Task compatibility depends on truth-table conflict, modulation distance and direction, and output polarity; identical Euclidean distances can yield opposite outcomes. Per-task activation removes these compatibility patterns, which arise from activation-task mismatch.

Task Compatibility in Modulatory Space

  • 100% convergence for OR+NAND versus 0% for OR+NOR despite identical Euclidean distance (d = 0.566), with truth-table conflict differing between the pairs.OR+NAND has 2/4 truth-table disagreements, while OR+NOR has 4/4.Source: Task Compatibility in Modulatory Space · p. 7
  • For complementary pairs, greater NT distance predicts better convergence: AND+NAND at d = 1.200 reaches 100%, versus OR+NOR at d = 0.566 reaching 0%.AND+NOR at d = 0.894 reaches 16.7%.Source: Task Compatibility in Modulatory Space · p. 7
  • Reorienting OR+NOR via mirror reflection restores 100% at generation 1, while changing NOR’s ACh polarity restores 100% at generation 0.Both rescues occur at identical Euclidean distance; polarity inversion makes complementary truth tables compatible via an output sign flip.Source: Task Compatibility in Modulatory Space · p. 7
  • All compatibility patterns vanish under per-task activation, indicating they arise from activation-task mismatch.The passage also identifies sufficient distance increase as a third way to rescue OR+NOR.Source: Task Compatibility in Modulatory Space · p. 7

Discussion

The discussion distinguishes an evolutionary search barrier from a representational limit, then describes activation diversity and depth as ways to address it. It connects these mechanisms to single-genotype behavioral diversity and notes boundaries in tasks, settings, and search methods.

Discussion

  • 100% success with Adam versus 0% under indirect evolutionary search and 76.7% under direct evolutionary search shows that the barrier depends on the optimizer, not representational capacity.Frozen tanh-trained weights yield 50% XOR accuracy even when sin is applied post-hoc, supporting a barrier during training dynamics rather than evaluation.Source: Discussion · pp. 7–8
  • Neuromodulation and activation diversity are jointly necessary: modulation tunes shared weights, while per-task activations supply task-appropriate computational primitives.A richer modulatory signal that selects activation functions could unify the two mechanisms.Source: Discussion · p. 8
  • Two-layer tanh also reaches 100%, but one-layer per-task activation is faster with fewer parameters; added depth eases evolutionary search rather than adding representational capacity.A single layer already represents the solution.Source: Discussion · p. 8
  • One indirectly encoded genotype can express qualitatively different behaviors depending on the task environment, a property relevant when an open-ended system’s task set is unknown in advance.The paper frames this as environment-dependent phenotypic expression at the network level.Source: Discussion · p. 8
  • For evolutionary systems, proposed paths include evolving activation diversity, differentiating neuron types during development, or composing specialized subnetworks with routing.The 2-task compatibility matrix is proposed as a space that quality-diversity algorithms could exploit.Source: Discussion · p. 8

Limitations

  • The evidence is bounded by Boolean-task scope, manually specified neuromodulator vectors, explicit task identity at evaluation, and the tested evolutionary algorithms.The encoding comparison is illustrative rather than strictly controlled; multi-objective, novelty- and diversity-promoting, and quality-diversity searches were not tested.Source: Limitations · pp. 8–9

Conclusion

Across more than 10,000 experiments, monotonic activations create an evolutionary search barrier, while task-matched activation functions enable multi-task success. The findings support treating computational primitives as evolvable traits so one genotype can express multiple behaviors.

Conclusion

  • 100% 5-task success (median 14 generations) with per-task activations eliminates the barrier, while monotonic activations impose a 75% ceiling on parity and asymmetric tasks.The oscillatory advantage generalizes across the activation class and is robust across a 40× range of modulation strengths.Source: Conclusion · p. 9
  • Computational primitives should be treated as evolvable traits, not fixed design choices; a single evolved topology can express different behaviors across tasks when its activation palette matches their computational demands.The conclusion recommends auditing the activation palette and exploiting output-level reuse via polarity inversion.Source: Conclusion · p. 9

Related papers

Loading 2610.00148v1…