Adaptive Directional Evolution Strategy

• Adaptive Directional Evolution Strategy (ADES) is a meta-evolutionary algorithm that combines directional mutation with step recording to enhance convergence on complex fitness landscapes.
• ADES employs a two-stage mutation process that first adapts mutation strengths and directional steps before applying phenotypic changes, ensuring robust directional exploitation.
• Empirical benchmarks show that ADES accelerates convergence in narrow, rotated valleys while maintaining linear computational complexity compared to full-covariance methods.

Adaptive Directional Evolution Strategy (ADES) is a meta-evolutionary optimization scheme combining directional mutation and a “step recording” mechanism within evolutionary programming algorithms. Its objective is to accelerate convergence, particularly on fitness landscapes characterized by long, narrow, and arbitrarily oriented valleys, while retaining robustness on multi-modal problems and linear computational scaling per individual (0803.3838).

1. Mathematical Structure of ADES

Each individual in the population at generation $t$ is represented as a triplet $(x, \sigma, k)$, where $x \in \mathbb{R}^n$ is the solution vector, $\sigma > 0$ is the omni-directional mutation strength, and $k \in \mathbb{R}^n$ is the directional “step” vector. The mutation operator comprises two stages:

A. Meta-mutation of $\sigma$ and $k$

• Draw $u \sim U(0,1)$ and set:

$\mu := \sigma + \frac{\|k\|}{10}$

$\sigma' := -\mu \ln(1-u)$

(where $(x, \sigma, k)$0 is sampled from an exponential distribution with mean $(x, \sigma, k)$1).

• Draw $(x, \sigma, k)$2.
• For each coordinate $(x, \sigma, k)$3,

$(x, \sigma, k)$4

B. Phenotypic Mutation of $(x, \sigma, k)$5

• For the ADES variant (directional + step recording):

$(x, \sigma, k)$6

• For pure directional mutation (no step recording):

$(x, \sigma, k)$7

The step-recording mechanism sets $(x, \sigma, k)$8 to $(x, \sigma, k)$9, and if $x \in \mathbb{R}^n$0 is selected, $x \in \mathbb{R}^n$1 is propagated, so the last successful directional step is recorded and re-used.

2. Step-Recording Mechanism

Step recording ensures $x \in \mathbb{R}^n$2—once computed from the meta-mutation stage—is exactly the step applied to $x \in \mathbb{R}^n$3 for the offspring. When an offspring is selected into the next generation, its complete genetic state $x \in \mathbb{R}^n$4 is retained. This enables persistent exploitation of productive search directions: once a useful direction emerges, the algorithm continues to bias evolution along this direction until it ceases to yield improved fitness.

No acceptance threshold or explicit memory is imposed; the only determinant for retention is standard survivor selection. All steps from selected individuals are recorded—not filtered.

3. Step-Size Adaptation and Mutational Dynamics

Step size adaptation is governed solely via updating $x \in \mathbb{R}^n$5 from the exponential law,

$x \in \mathbb{R}^n$6

This coupling means that large directional mutations (large $x \in \mathbb{R}^n$7) automatically increase the global step size, avoiding premature convergence or “locking in” to suboptimal scales.

The update for $x \in \mathbb{R}^n$8 per coordinate,

$x \in \mathbb{R}^n$9

introduces correlated meta-mutation, inducing spatial autocorrelation without constructing a full covariance matrix.

These mechanisms yield a self-similar, orientation-invariant adaptation of both omni-directional and directional step magnitudes.

4. Algorithmic Structure and Complexity

The evolutionary algorithm proceeds as follows:

1. Initialization: For each individual $\sigma > 0$0, $\sigma > 0$1 is randomly set within the search domain, $\sigma > 0$2 is initialized (e.g., 1.0), and $\sigma > 0$3 is set to the zero vector.
2. Reproduction for each parent $\sigma > 0$4 and offspring $\sigma > 0$5:
   • Compute $\sigma > 0$6.
   • Sample $\sigma > 0$7, $\sigma > 0$8; $\sigma > 0$9.
   • For all $k \in \mathbb{R}^n$0, $k \in \mathbb{R}^n$1, $k \in \mathbb{R}^n$2.
   • Form offspring $k \in \mathbb{R}^n$3.
3. Selection: Evaluate fitness for all offspring; select the best $k \in \mathbb{R}^n$4 individuals for the next generation.

Per individual, ADES stores $k \in \mathbb{R}^n$5 real parameters and requires $k \in \mathbb{R}^n$6 operations per mutation, leading to a total generational complexity of $k \in \mathbb{R}^n$7. In comparison, full-covariance evolution strategies require $k \in \mathbb{R}^n$8 storage and computation, which is prohibitive for high-dimensional problems (0803.3838).

5. Rotational Invariance and Sampling Geometry

The combination of an isotropic Gaussian, $k \in \mathbb{R}^n$9, and a directional bias $\sigma$0 produces an ellipsoidal Gaussian whose principal axis is aligned with $\sigma$1 but not with any coordinate axis. Under any orthonormal rotation $\sigma$2, the mapping $\sigma$3 leaves the offspring distribution invariant. Thus, ADES is fully rotationally invariant, in contrast to self-adaptive schemes maintaining coordinate-wise variances $\sigma$4, which break this symmetry and may underperform when the natural orientation of the fitness landscape is not axis-aligned (0803.3838).

6. Empirical Performance and Benchmarks

Empirical tests reported in the primary study include three canonical optimization problems (up to $\sigma$5):

• F1: $\sigma$6 (symmetric quadratic)
• F6: Bohachevsky function (multi-modal)
• F9: $\sigma$7 (narrow rotated valley)

Four algorithmic variants were benchmarked: A) Classical meta-evolution (MEP) B) MEP + recorded step (MEP+RS) C) MEP + directional only (MEP+DM) D) MEP + RS + DM (ADES)

Performance Regimes:

• On F1, all variants converged within $\sigma$810–20 generations to $\sigma$9 accuracy; MEP slightly outperformed in speed but differences were minor.
• On F6, only MEP+DM lagged, while ADES matched the robustness of alternatives.
• On F9, ADES (MEP+RS+DM) alone ensured robust, rapid convergence; other variants stagnated due to inability to adapt mutation orientation.

Selected median error results after 50 generations:

Function	MEP	MEP+RS	MEP+RS+DM (ADES)
F1	$k$0	$k$1	$k$2
F6	0	0	0

On the F9 narrow valley, only ADES reached errors $k$3 within 50 generations, highlighting its ability to align search efficiently along arbitrarily oriented valleys (0803.3838).

7. Relation to Other Evolutionary Strategies

ADES provides directional learning and covariance-structure exploitation without the memory or computational cost of full-covariance matrix adaptation. Methods maintaining one variance per coordinate (as in standard self-adaptive schemes) are not rotationally invariant and can be orders of magnitude slower on landscapes where the axis of the valley is rotated relative to the coordinate axes. ADES achieves performance comparable to standard evolutionary programming methods on multi-modal problems and demonstrates substantial acceleration on rotated, high-aspect-ratio landscapes, with linear space and time complexity (0803.3838).