Sparse Cosine Optimized Policy Evolution (SCOPE)

• SCOPE is a method that uses a 2D discrete cosine transform to compress high-dimensional sensor data, reducing input size by 98% without sacrificing key temporal–spatial features.
• It enables evolutionary optimization by operating on a sparse representation of data, leading to a 20% improvement in locomotion performance in hexapod experiments.
• The approach integrates a linear policy architecture with a steady-state genetic algorithm, effectively mitigating the curse of dimensionality in policy evolution.

Sparse Cosine Optimized Policy Evolution (SCOPE) is an approach designed to address the scalability limitations of evolutionary algorithms (EAs) in high-dimensional policy search tasks, particularly as applied to robotic gait generation. By leveraging the discrete cosine transform (DCT) to sparsify and compress high-dimensional state observations, SCOPE enables more efficient evolutionary optimization and significant reductions in policy parameterization, without sacrificing the representation of critical temporal–spatial input features (O'Connor et al., 17 Jul 2025).

1. Motivation and Conceptual Overview

As controller input dimensionality grows, the parameter space for evolutionary policy search expands exponentially, hampering convergence and ultimately degrading the performance of EAs. This is particularly problematic in domains such as hexapod locomotion, where a rich time-series of high-dimensional sensorimotor data is required to encode adaptive gaits. SCOPE addresses this “curse of dimensionality” by reformulating the EA input pipeline:

• The raw observation matrix (e.g., a time-series of motor sensor values) is transformed using a two-dimensional, type-II discrete cosine transform (2D DCT).
• Only the lowest-frequency (highest-energy) cosine coefficients are retained, forming a small block that captures the most informative components of the signal.
• Policy evolution is then performed over this sparse, compressed input, yielding a drastic reduction in the number of parameters and a commensurate improvement in sample efficiency and final policy efficacy.

By concentrating signal energy into a reduced subset of DCT features, SCOPE achieved a 98% reduction in input size (from 2,700 to 54 dimensions) and a 20% increase in mean efficacy on the target locomotion task (O'Connor et al., 17 Jul 2025).

2. Mathematical Formulation: 2D Type-II DCT and Input Transformation

Let $M \in \mathbb{R}^{m \times n}$ denote the input matrix, where $m=6$ corresponds to the number of robot legs and $n=450$ aggregates 50 time steps of 9 features per step (position, velocity, acceleration for each joint). SCOPE applies the standard separable, orthonormal type-II DCT as follows.

For 1D $x \in \mathbb{R}^N$:

$$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$

where $\alpha_k = \sqrt{1/N}$ if $k=0$ and $\sqrt{2/N}$ otherwise.

The extension to 2D is given by:

$C = D_2(M) = A_m \cdot M \cdot A_n^\top$

or elementwise:

$$C_{u,v} = \sum_{i=0}^{m-1} \sum_{j=0}^{n-1} M_{i,j} \cos\left[\frac{\pi}{m}(i+\frac{1}{2})u\right] \cos\left[\frac{\pi}{n}(j+\frac{1}{2})v\right]$$

for $m=6$0, $m=6$1. Here, $m=6$2 is the coefficient at frequency $m=6$3 and $m=6$4, $m=6$5 are the DCT basis matrices.

The DCT's energy compaction property guarantees that most of the signal's $m=6$6 energy is concentrated in low-frequency coefficients ($m=6$7 small).

3. DCT Coefficient Truncation and Dimensionality Reduction

After computing $m=6$8, a block truncation is performed to extract the $m=6$9 lowest-frequency coefficients:

• Selected integers $n=450$0, $n=450$1 (with $n=450$2, $n=450$3 in the reference experiment).
• The truncated matrix is $n=450$4.

This direct truncation preserves the most significant features along both spatial and temporal axes while achieving dramatic input compression:

$n=450$5

For the hexapod scenario, this reduces the input from $n=450$6 to $n=450$7 dimensions (a $n=450$8 reduction).

4. Policy Architecture and Evolutionary Search Integration

The $n=450$9 DCT coefficients are vectorized to form $x \in \mathbb{R}^N$0, serving as input for the policy. The policy mapping is purely linear:

$x \in \mathbb{R}^N$1

where $x \in \mathbb{R}^N$2 and $x \in \mathbb{R}^N$3, generating 18 outputs (grouped as $x \in \mathbb{R}^N$4 for 6 legs × 3 joints).

Each motor is then actuated using a central pattern generator (CPG):

$x \in \mathbb{R}^N$5

with constraints on joint transitions to ensure smoothness. The SSGA genotype comprises the flattened $x \in \mathbb{R}^N$6 and $x \in \mathbb{R}^N$7, totaling 108 free parameters.

The evolutionary optimization employs a steady-state genetic algorithm (SSGA) with standard tournament selection, crossover, and Gaussian mutation. At each episode boundary (every 3 s within a 15 s run), the most recent sensor history is transformed via DCT truncation, and the policy is evaluated according to the Euclidean distance covered by the robot.

5. Implementation Details: Algorithmic and Experimental Setup

The following pseudocode outlines the core SCOPE-SSGA loop, with population size $x \in \mathbb{R}^N$8, generations $x \in \mathbb{R}^N$9, and tournament parameters as per the reference:

$\alpha_k = \sqrt{1/N}$0

Experiments are conducted in Webots with a mantis-inspired hexapod (6 legs × 3 joints), using position, velocity, and acceleration readings per joint to form each time slice. Each policy is evaluated over 500 independent runs (O'Connor et al., 17 Jul 2025).

6. Experimental Results: Compression, Efficacy, and Convergence

SCOPE yields substantial quantitative improvements relative to uncompressed baselines. The following summarizes key metrics:

Method	Input Dim	Params	Mean Fitness
Baseline	2,700	5,400	11.880
SCOPE	54	108	14.242

• SCOPE compresses the policy input from $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$0 to $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$1 dimensions, and parameter count from $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$2 to $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$3 (98% fewer).
• Mean fitness improves by 20% compared to the baseline, as measured by distance traveled, with this difference statistically significant ($$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$4, $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$5, Mann–Whitney U test).
• Convergence curves indicate that the performance advantage for SCOPE is maintained throughout 5,000 generations.

The efficacy increase is directly attributable to the reduction in search space dimensionality, which accelerates evolutionary convergence without sacrificing the representation of key time-varying features (O'Connor et al., 17 Jul 2025).

7. Applicability, Limitations, and Extensions

SCOPE makes no domain-specific assumptions: any $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$6 input matrix can be DCT-compressed to $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$7, provided $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$8 and $$X_k = \alpha_k \sum_{i=0}^{N-1} x_i \cdot \cos\left[\frac{\pi}{N}\left(i+\frac{1}{2}\right)k\right]$$9. The truncation shape may be tailored or permuted for different downstream models, including neural networks and attention mechanisms.

Potential limitations arise if critical high-frequency information (e.g., sudden events or noise signatures) is lost through low-frequency DCT truncation. Adjustment via percentile thresholding or more adaptive sparsification could be required in such scenarios.

Extensions include application to high-dimensional perceptual tasks (such as visual or Atari-like environments with significant background noise), as well as integration with alternative evolutionary strategies (e.g., CMA-ES, MAP-Elites) or hybrid pipelines utilizing compressed DCT features as input to deeper models.

Overall, SCOPE establishes a straightforward linear compression method that enables evolutionary algorithms to effectively operate on high-dimensional time-series data by extracting the most salient low-frequency temporal–spatial features, facilitating both faster convergence and improved control performance (O'Connor et al., 17 Jul 2025).