🌀 nonlinear sciences

Directing Open-Ended Evolution in Artificial Life via Multi-Scale Path Divergence

This paper introduces Multi-Scale Path Divergence (MSPD), a renormalization-group-inspired scalar metric that quantifies temporal multiscale heterogeneity to both guide open-ended evolution in artificial life systems and provide a principled, interpretable bridge to physical theories of complexity, outperforming black-box neural network approaches across diverse substrates.

Original authors: Mikhail Akhtyrchenko, Mikhail I. Katsnelson, Andrey Ustyuzhanin

Published 2026-06-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mikhail Akhtyrchenko, Mikhail I. Katsnelson, Andrey Ustyuzhanin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a bustling city from a helicopter. You see cars moving, people walking, and lights changing. Some parts of the city look chaotic (like a traffic jam), some look boring (like an empty parking lot), and some look incredibly complex (like a festival with parades, markets, and changing weather).

For a long time, scientists trying to create "artificial life" (computer simulations that act like living things) have struggled with a simple question: How do we tell the difference between just "moving around" and actually being "complex"?

Most current methods use a "black box" computer brain (a neural network) to guess. It says, "This looks interesting!" but it can't explain why, and it doesn't connect to the laws of physics.

This paper introduces a new tool called MSPD (Multi-Scale Path Divergence). Think of MSPD not as a black box, but as a smart, transparent ruler that measures how "surprising" and "organized" a system is over time.

Here is how it works, broken down into simple concepts:

1. The Core Idea: Watching the "Rules" Change

Imagine you are watching a dance troupe.

Boring Scenario: Everyone does the exact same move at the exact same time. (Low complexity).
Chaotic Scenario: Everyone is flailing randomly. (Low complexity, just noise).
Complex Scenario: The dancers start in a line, then split into two groups doing different dances, then those groups split again, and the timing of these splits is perfectly coordinated.

MSPD doesn't just count how many dancers are moving. It looks at the local rules of the dance. It asks: "If I look at one dancer, what are they likely to do next? If I look at another dancer, is their rule different? And how do these rules change as time goes on?"

If the rules for different parts of the system stay the same forever, the score is low. If the rules change in a coordinated, multi-layered way (like the dance troupe splitting and re-forming), the score goes up.

2. The "Frustration" Metaphor

The paper connects this to a physics concept called "Frustration."

Imagine a group of friends trying to decide where to eat.

Easy: Everyone agrees on Pizza. (Low frustration, low complexity).
Impossible: Everyone wants a different place, and they are all stuck in a loop of arguing. (High frustration).

In physics, "frustrated" systems are the ones that can't settle down into a single, boring state. They are constantly juggling conflicting demands. The paper argues that real life is "frustrated." It's never perfectly settled; it's always balancing conflicting forces.

MSPD measures this "frustration" without trying to. It simply measures how much the system's behavior at a small scale (one dancer) differs from its behavior at a larger scale (the whole troupe). The more these scales disagree and interact in complex ways, the higher the score.

3. The Experiment: Teaching Computers to Evolve

The researchers used MSPD as a "fitness score" to guide artificial life simulations. They told the computer: "Evolve your rules to get the highest MSPD score possible."

They tested this on three different types of digital worlds:

Cellular Automata: Like a grid of pixels (think Game of Life) where cells are either on or off.
Flow-Lenia: A fluid-like world where "mass" flows around, like water or smoke.
Particle Life++: A world of individual particles (like atoms) that push and pull each other based on neural networks.

The Results:

The "Smart" vs. The "Random": When they let the computer evolve using MSPD, the resulting systems were significantly more complex than systems created by random chance.
The "Instability" Test: They found that the high-scoring systems were "fragile" in a good way. If you tweaked them slightly (like changing one pixel or nudging one particle), the whole system would change dramatically. This is a hallmark of living things: they are sensitive to their environment.
The "Frustration" Check: The high-scoring systems showed exactly the kind of "scale-dependent frustration" that physicists say defines biological complexity. They were juggling conflicting rules at different levels of organization.

4. Why This Matters (According to the Paper)

The paper claims that MSPD is a universal translator between Artificial Life and Physics.

No Black Boxes: Unlike other methods, MSPD is a clear, mathematical formula. You can look at it and understand exactly what it's measuring.
It Works Everywhere: It worked on grid-based pixels, fluid flows, and particle systems.
It Finds "Life": It successfully identified systems that behave like living things (unstable, complex, multi-layered) without being explicitly programmed to look for "life."

Summary Analogy

Think of MSPD as a music critic for computer simulations.

Old methods were like a critic who just says, "This song sounds cool," but can't tell you why.
MSPD is a critic who listens to the rhythm, the harmony, and how the melody changes over time. It knows that a song where every instrument plays the same note forever is boring. It knows that a song where instruments play random noise is just static.
But it loves a song where the instruments start simple, then split into different sections, then come back together in a way that feels both structured and surprising.

The paper proves that when you teach a computer to write "songs" that get a high score from this critic, the computer accidentally discovers the same complex, "frustrated" behaviors that we see in real biology.

Technical Summary: Directing Open-Ended Evolution in Artificial Life via Multi-Scale Path Divergence

1. Problem Statement

The field of Artificial Life (ALife) aims to achieve Open-Ended Evolution (OEE)—the sustained generation of novel forms and behaviors. However, current approaches to driving OEE rely heavily on uninterpretable, black-box metrics derived from neural networks (e.g., novelty search, quality-diversity archives). These metrics lack a principled connection to physical theories of complexity, leaving life-like systems disconnected from the underlying physics that governs real biological complexity.

While Multi-Scale Structural Complexity (MSSC) offers a physics-grounded approach for static spatial patterns based on renormalization group (RG) theory, it is inherently static. Real biological systems are dynamic. The central challenge addressed by this paper is defining a formal, explicit, and computable scalar that:

Simultaneously optimizes for complexity and reveals its physical basis.
Extends the concept of spatial multiscale complexity into the temporal domain.
Avoids black-box mechanisms, providing a transparent link to physical theories such as frustration (the inability of a system to settle into a single global equilibrium due to competing interactions).

2. Methodology: Multi-Scale Path Divergence (MSPD)

The authors introduce Multi-Scale Path Divergence (MSPD), denoted as $D_P$ , a renormalization-group-inspired scalar metric. MSPD quantifies the temporal multiscale organization of heterogeneity in local transition laws.

2.1 Theoretical Framework

The method is built upon a configuration model where a system's state is a map $x: \mathcal{I} \to E$ (local degrees of freedom to local states).

Local Transition Laws: For a current configuration $x$ , the system defines a transition law $T^i_{\Delta t}(x)$ for each local degree of freedom $i$ . This represents the probability distribution of the state change at $i$ given the global configuration.
State-Level Dynamical Heterogeneity ( $\mathcal{H}_{\Delta t}$ ): This measures the pairwise dissimilarity between the transition laws of different local components within a single state. If all local components behave identically, heterogeneity is zero. If different regions exhibit distinct dynamical rules, heterogeneity increases.
Pathwise MSPD: Unlike static heterogeneity, MSPD measures how this heterogeneity trace evolves over time across different temporal scales. It treats the heterogeneity trace as a function and computes its sensitivity to temporal coarse-graining.
- The metric is defined as the integral of the squared logarithmic scale derivative of the heterogeneity trace, normalized by the signal power at that scale.
- Mathematically, it captures systems where the dynamics expressed at different temporal resolutions differ significantly (scale-dependent organization), distinguishing them from stationary or purely random noise.

2.2 Computational Estimator

Since exact transition laws are unobservable in simulations, the authors propose a windowed finite-resolution estimator:

Trajectory Tracking: The system tracks local carriers (particles, lattice sites, or mass probes) over time.
Windowed Transition Laws: Local transition laws are estimated empirically over finite time windows $I_k$ and lags $\tau$ .
Pairwise Dissimilarity ( $\Delta H$ ): The Wasserstein-1 distance (approximated via Sliced Wasserstein, $SW_1$ ) is used to measure the divergence between the empirical transition laws of different carriers.
Aggregation: The estimator aggregates these pairwise dissimilarities across a grid of temporal scales to compute the final MSPD score.
Consistency: The paper provides propositions proving that the empirical estimator converges to the theoretical pathwise functional under conditions of local stationarity and sufficient sampling.

2.3 Application to Substrates

The protocol is applied to three distinct substrates, adapting the definition of "local carriers" and "transition laws" for each:

Life-like Cellular Automata (CA): Local carriers are lattice sites; transition laws are categorical (before/after neighborhood states).
Flow-Lenia: A mass-conserving continuous CA. Local carriers are passive mass tracers; transition laws are empirical laws of finite-time displacements.
Particle Life++: A neural-network-mediated particle system. Local carriers are particles; transition laws are empirical laws of particle displacements derived from neural interaction weights.

3. Key Contributions

The paper makes four primary contributions:

Definition of MSPD: Introduces MSPD as a temporal extension of spatial MSSC, defined as a functional of the realized trajectory via local transition laws, with a consistent windowed finite-resolution estimator.
Dual Role: Demonstrates that MSPD functions effectively both as a gradient-free fitness function for evolutionary optimization and as a post-hoc analytical lens to expose the local transition laws of any simulation.
Separation of Signal from Noise: Shows that MSPD-optimized parameters produce higher held-out complexity scores than matched random parameters. Crucially, high- $\Delta H$ states (high heterogeneity) are linked to higher instability under external interventions, indicating the metric tracks underlying dynamics rather than noise.
Physical Grounding via Frustration: Establishes that higher MSPD corresponds to stronger scale-dependent frustration. Optimized systems exhibit larger differences between dynamics expressed at different spatial extents compared to unconstrained rollouts. This links MSPD directly to the frustration criterion of biological complexity (Vanchurin et al., Wolf et al.) without the metric being explicitly optimized for frustration.
Generalizability: Validates that the protocol transfers across qualitatively different substrates (discrete CA, continuous mass fields, and neural particle systems).

4. Experimental Results

The authors validate MSPD through a series of claims (N0, C1, C2, C5):

N0 (Synthetic Calibration): MSPD correctly distinguishes between static noise, homogeneous motion, and temporally organized complexity.
- Result: Static particles and coherent moving blobs (S0, S3) yield near-zero MSPD. Systems with temporally extended transitions or split events (S6, S8) yield high MSPD. This confirms MSPD measures temporal organization of heterogeneity, not just raw motion or visual novelty.
C1 (Optimization vs. Random Controls):
- Particle Life++: Optimized parameters produced significantly higher MSPD scores than matched random controls (statistically confirmed).
- Life-like CA: Exhaustive sweeps identified rules with higher MSPD than random rules.
- Flow-Lenia: Optimized checkpoints showed positive contrast in 6/9 groups, though statistical significance was not reached under the current finite-budget protocol (attributed to rollout noise).
C2 (Instability to Intervention):
- Result: States with higher present-time local transition heterogeneity ( $\Delta H$ ) exhibited greater future divergence when subjected to small perturbations (bit flips for CA, Gaussian noise for Flow-Lenia). This confirms that high-MSPD states are dynamically unstable and sensitive to intervention.
C5 (Scale-Dependent Frustration):
- Result: High-MSPD systems showed larger differences between constrained and unconstrained rollouts compared to random controls. This demonstrates that the metric captures the "frustration" signature (non-ergodicity and rugged landscapes) characteristic of biological complexity, even though the optimization objective was purely MSPD.

5. Significance and Claims

The paper claims that MSPD provides a principled bridge between the engineering of ALife systems and the physics of complexity.

Beyond Black Boxes: Unlike neural-network-based drivers, MSPD is an explicit formula. It does not require training or hand-tuned descriptors.
Physical Interpretability: By linking high complexity to scale-dependent frustration, MSPD aligns with physical theories (spin-glass analogies) that view life as a form of frustrated matter. It suggests that biological complexity arises from the system's inability to settle into a single equilibrium, a property that MSPD quantifies directly.
Unified Framework: The ability to apply the same metric across discrete, continuous, and neural substrates suggests a universal approach to measuring and directing open-ended evolution based on the temporal organization of local dynamics.

The authors maintain a modest tone, acknowledging limitations such as the sensitivity of the metric to window/lag selection and the computational cost of high-resolution trajectory tracking. They position MSPD not as a universal definition of complexity for all processes, but as a specific, measurable aspect of dynamical complexity for systems where local transition laws are meaningful.