🤖 AI

RDEx-CASK: Cauchy Mutation, Archive, and Stagnation Kick for RDEx-CSOP

The paper introduces RDEx-CASK, an enhanced version of the RDEx-CSOP algorithm that integrates a truncated Cauchy mutation, a feasible-only JADE-style archive, and a stagnation-triggered local override mechanism to improve feasibility-aware quality and time-to-target performance on CEC constrained optimization problems.

Original authors: Dikshant, Dikshit Chauhan, Chen Hao, Anupam Trivedi, Harikumar Kandath, Senthilnath Jayavelu

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dikshant, Dikshit Chauhan, Chen Hao, Anupam Trivedi, Harikumar Kandath, Senthilnath Jayavelu

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 leading a team of treasure hunters (the algorithm) searching for a hidden gold mine (the best solution) in a vast, dangerous jungle (the complex problem space). The jungle is full of traps (constraints) that can disqualify you, and there are many dead-end caves (local optima) that look like the real mine but aren't.

The paper introduces an upgraded version of a specific treasure-hunting strategy called RDEx-CASK. Think of the original strategy (RDEx) as a very good, experienced guide. The new version, RDEx-CASK, keeps the guide's core wisdom but adds three specific "emergency tools" to help the team when they get stuck or start moving too slowly.

Here is how the three new tools work, using simple analogies:

1. The "Wild Card" Jump (Cauchy Mutation)

The Problem: Sometimes the team gets stuck in a small, flat area of the jungle. They are taking tiny, cautious steps, but they aren't finding anything new.
The Old Way: The guide used to take two steps that were perfectly synchronized, like a soldier marching. If the first step was small, the second was small too.
The New Way (CASK): The guide now takes a second step that is independent and unpredictable. Most of the time, it's a normal step. But occasionally, this "Wild Card" step is a massive leap (like a kangaroo jump).
Why it helps: These rare, huge leaps allow the team to suddenly jump over a small hill or a dead-end cave they were stuck in, without making the whole team reckless. It's like having a trampoline that only pops up once in a while to help you escape a rut.

2. The "Safe Vault" (JADE-Style Archive)

The Problem: The team needs to remember where they've been to avoid going in circles, but they also don't want to remember bad places.
The Old Way: The team didn't have a special place to store "almost good" ideas.
The New Way (CASK): The guide adds a small, secure Safe Vault (an archive).

The Rule: You can only put a map into the vault if you found a spot that was safe (feasible). If you found a spot that was safe but slightly better than your current spot, you swap your current spot with the new one, and the old spot goes into the vault.
The Usage: When the team is stuck, the guide occasionally pulls a map out of the vault to try a different path.
Why it helps: It keeps a collection of "safe but forgotten" ideas. If the team is wandering aimlessly, they can look back at these safe spots to find a new direction, but they never look at the dangerous, trap-filled spots that were rejected earlier.

3. The "Wake-Up Call" (Stagnation Kick)

The Problem: Sometimes, a specific hunter in the team stops finding better spots for a long time (180 days, in the paper's terms). They are just walking in place.
The Old Way: The team would just keep waiting for that hunter to figure it out on their own.
The New Way (CASK): The guide has a stopwatch for every hunter. If a hunter hasn't improved in 180 steps, the guide hits the Wake-Up Call:

Pull to the Leader: The guide forces that hunter to stop wandering and start walking directly toward the current best location found by the whole team (the Global Best).
Force the Vault: The guide forces that hunter to use the "Safe Vault" much more often (65% of the time) to try new, safe paths.
Go Wild: If the whole team is failing to find any new safe spots, the guide tells that hunter to take huge, risky steps (high crossover rate) to break the deadlock.
Why it helps: It prevents individual team members from wasting time. If they are stuck, the guide intervenes with a specific, temporary boost to get them moving again.

The Result: Faster, Not Necessarily "Perfect"

The paper tested this new strategy on 28 different "jungle maps" (benchmark problems).

Speed: RDEx-CASK found good solutions much faster than the original guide in most cases. It reached the "finish line" sooner.
Quality: In some very tricky, complex jungles, the original guide might have found a slightly better final spot if given infinite time. However, RDEx-CASK is very competitive and often finds a spot that is just as good, but in less time.
Safety: Crucially, the new strategy never lost its ability to find safe spots (feasible solutions) where other strategies failed.

In summary: RDEx-CASK is like upgrading a treasure hunting team with a trampoline for big jumps, a vault for safe memories, and a strict manager who wakes up lazy team members. The result is a team that finds the treasure much faster without getting lost or falling into traps.

Technical Summary: RDEx-CASK

Problem Statement

The paper addresses the challenge of solving Constrained Optimization Problems (COPs) using Differential Evolution (DE), specifically targeting the limitations of the existing RDEx-CSOP algorithm. While RDEx-CSOP achieved top rankings on the CEC CSOP track due to its feasibility-aware ranking and fast threshold attainment, the authors observed two specific failure modes in the standard branch of the algorithm:

Parameter Rigidity: Certain adaptation parameters were tuned for an earlier benchmark cycle and lacked robustness for the current problem set.
Late-Stage Stagnation: On many problems, the "front population" (the subset of individuals driving the search) enters long plateaus of no improvement during the second half of the computational budget. During these periods, individual targets cease to contribute accepted trials, effectively stalling the search process.

The goal is to extend RDEx-CSOP to mitigate these stagnation issues and improve time-to-target metrics without compromising the algorithm's feasibility-aware quality or disrupting its successful Exploitation-Biased (EB) branch.

Methodology: RDEx-CASK

The proposed variant, RDEx-CASK (Cauchy, Archive, Stagnation Kick), introduces three structural modifications and several parameter refinements. Crucially, all changes are confined to the standard branch of the algorithm; the EB branch, constraint handling rules, population reduction schedule, and success-history bookkeeping remain untouched.

1. Independent Cauchy Second Scale Factor

In the baseline RDEx, the second scale factor ( $F_2$ ) in the mutation vector is identical to the first ( $F$ ), creating perfect correlation between difference components. RDEx-CASK decouples these by sampling $F_2$ independently from a truncated Cauchy distribution ( $Cauchy(\mu_F, 0.1) \cap [0, 1]$ ).

Rationale: The heavy tails of the Cauchy distribution allow for occasional large jumps to escape local plateaus without permanently inflating the search variance, a behavior not replicated by Normal distributions in ablation tests.

2. JADE-Style External Archive

A circular archive ( $A$ ) with a capacity of 50 is added to the standard branch.

Mechanism: When a trial vector replaces a target, the displaced parent is pushed to the archive only if the displaced parent was feasible (violation $\phi \le \epsilon$ ).
Sampling: The standard branch samples a vector from this archive with probability $p_{arch} = |A| / (|A| + |P|)$ .
Rationale: This introduces diversity from past successful solutions. The "feasibility gate" is critical; pushing infeasible parents into the archive was found to pull trials off the feasibility manifold, degrading performance.

3. Per-Individual Stagnation Kick

A diagnostic counter ( $\sigma_i$ ) tracks the number of generations an individual $i$ goes without an accepted trial. If $\sigma_i \ge 180$ (stagnation threshold), three local overrides are triggered for that specific individual for one generation:

S1 (Global-Best Substitution): The $p$ -best donor in the mutation formula is replaced by the running global best ( $x^*$ ). Notably, this pull is applied even if $x^*$ is infeasible, as the authors found that pulling toward an infeasible best accelerates feasibility attainment in infeasible regimes.
S2 (Archive Probability Floor): The probability of sampling from the archive is lifted to a floor of 0.65, ensuring a higher injection of historical diversity for the stagnated individual.
S3 (Crossover Saturation): The crossover rate ($CR$) is saturated to 0.95, but only if the overall population success rate is below 0.10. This double-gate prevents premature saturation during normal convergence phases.

4. Parameter Refinements

Memory Length: Increased from $H=5$ to $H=10$ to match the LSHADE family.
Success-Rate Exponent: The exponent for the mean of the Gaussian distribution for $F$ was changed from $1/3$ to $2/5$ to maintain directional pressure in early phases.
Local Perturbation: Disabled the Cauchy perturbation on non-crossover coordinates to preserve target values.
EB Rate Initialization: Lowered the initial Exploitation-Biased rate from 0.7 to 0.5.
Smoothed EB Update: Replaced hard resets of the EB rate with a smoothed update rule.

Experimental Results

The algorithm was evaluated on the CEC CSOP suite (28 constrained problems, dimension $D=30$ , MaxFE = $20000 \times D$ ) over 25 independent runs. It was compared against the baseline RDEx, UDE-III, and CL-SRDE.

Feasibility-Aware Quality ( $Q_p$ ): RDEx-CASK is competitive with the other algorithms. It improves upon RDEx on problems F7, F11, F17, F21, F26, and F28. However, on some multimodal landscapes (e.g., F4, F12, F13), the aggressive stagnation kicks (global-best pull and high CR) resulted in slightly lower final accuracy compared to the baseline.
Time-to-Target (TTT): RDEx-CASK demonstrates a significant improvement in speed. It reaches the median final-quality target much faster than RDEx, UDE-III, and CL-SRDE on the majority of problems.
Robustness: The algorithm maintains feasibility on all 28 problems where any tested algorithm reaches feasibility.

Significance and Claims

The paper claims that RDEx-CASK successfully targets the specific issue of late-stage stagnation in constrained differential evolution. By introducing a "local, gated, and bounded" stagnation kick, the authors achieve a trade-off: a small potential loss in final objective accuracy on specific multimodal problems in exchange for a significant improvement in the speed component of the U-score (time-to-target).

The authors emphasize that the structural changes are modular and isolated to the standard branch, preserving the integrity of the successful EB branch and constraint handling mechanisms of the original RDEx-CSOP. The results suggest that decoupling scale factors, utilizing a feasible-only archive, and applying conditional stagnation overrides are effective strategies for maintaining search momentum in the latter half of the optimization budget.