Lower Power Genetic Algorithm

⚠️ Curated Limitations & Negative Results

• The printed d_avg equation sums only distances through the user-specified d_max and divides by N squared; it is not the unrestricted average distance when longer paths exist. The paper separately reports N_over. Its adjacent-node example uses distance one, while other prose describes intermediate nodes. The distance convention and conversion of the minimized distance score into ranked fitness are not fully specified; neither an equation correction nor a particular fitness transformation is imposed here.
• Swap mutation preserves label multiplicities and its checks prevent self-assignment/protected-position changes. Those constraints alone do not imply a connected processor network. The selected source does not fully specify disconnected-candidate distance handling or an additional connectivity test. No claim is made that the authors’ implementation actually generated such candidates.
• The production-table rule asks for k equal groups, but the reported population of 20 with k=3 requires an unstated boundary/rounding convention. Parent quotas, survivor ties, shortest-path ties and retention-percentage rounding do not uniquely specify an executable algorithm. Random choices are not automatically uniform.
• The retry-until-legal mutation loop requires a permissible pair. A restrictive schema can leave no legal pair or only no-op swaps; no fallback or finite retry bound is given. The fully writable first phase avoids protection-induced rejection, but it does not by itself guarantee a nontrivial mutation; no failure is asserted for the reported retention percentage. Protected positions restrict reachable configurations and can exclude better ones. Fixed generation budgets do not certify a local or global optimum.
• CPA counts selected shortest-path usage under its own tie conventions, not externally measured packet traffic. Protecting high-hit links and removing a lowest-hit link do not prove optimal future configurations or minimum deterioration of either objective. Explicit enumeration of all shortest paths can be expensive; the Low/Lower Power name does not provide a general complexity or speed bound.
• The paper’s finite 32–40-processor comparisons do not support universal superiority. Table 1 includes n=36, where LPGA(CPA) has d_avg 2.39 versus GAcSP(HC) 2.38; Table 2 also has rows where CPA is worse. Mean distance and time values and best-observed overshoot counts use different aggregation, and AMP execution times are unavailable. The reported results have not been independently reproduced here.
• This account is bounded to the inspected 1996 eight-page author manuscript, which uses both Lower Power and Low Power for LPGA and motivates Power by simplicity and computational cost, not a processor-energy objective. The 1997 journal paper is bibliographically identified, but its full text was not inspected; retained publisher access attempts returned HTTP 403. Later GGA and thesis passages supply lineage context, not exact version equivalence or a reason to import adaptive GGA penalties. The original AMP and GAcSP comparator methods were not inspected in this research.