Dingo Optimization Algorithm (DOA / DOAm)

A population metaheuristic modelled on dingo pack behaviour, and andrey-dik‘s modified variant DOAm (MQL5 article, 2026-07-24) which tops his own ~45-algorithm suite at 73.63% — displacing (μ+λ)-ES (72.18%) as benchmark leader.

The base algorithm — four behaviours

DOA switches between four strategies rather than running one update rule:

• Group attack — coordinated movement using several pack members.
• Persecution — an individual chase toward selected prey.
• Scavenging — solo foraging, movement without a target.
• Survival procedure — an emergency tactic for agents whose fitness is falling.

Stripped of the metaphor: a mode-switching optimizer, where a per-agent condition (fitness rank, a probability threshold) selects between a social move, a greedy move, a random move, and a rescue move. The nearest relative in the corpus is GWO — also a canid pack story — but GWO applies one blended pull toward three leaders where DOA branches into separate behaviours.

What DOAm changes

Three small edits, and their smallness is the point:

1. Sign flip in group attack — add to the best solution instead of subtracting from it, i.e. “movement towards a leader instead of moving from it.” The attack vector is also divided by na × coords rather than na.
2. Survival threshold 0.3 → 0.01 — so only the genuinely worst agents trigger the survival branch, which for most iterations switches the mode off.
3. Absolute value dropped from the scavenging formula — negative coordinates become reachable, giving “freer movement in the search space.”

A sign error, a threshold, and a stray abs(). None of these touch the metaphor or add a mechanism; they correct the direction and gating of moves the original already had.

Benchmark result and its caveats

Composite 6.627 / 9.0 = 73.63% over Hilly / Forest / Megacity at 5, 25 and 500 dimensions (10,000 evaluations, 50 runs). Per-function averages: Hilly 0.454, Forest 0.912, Megacity 0.848 — strong on the rugged and the discrete, comparatively weak on the smooth. Runners-up in Dik’s table: ANS 68.15%, CLA 67.86%, AMOm 66.52%; random walk sits at 26.09%.

Dik names the weaknesses himself: high run-to-run variance (needs multiple independent runs), a tendency to stick, especially on smooth functions, and modest convergence on simple landscapes. His recommendation is to use DOAm as an exploration front-end and hand off to a precise optimizer for refinement — which is exploration-vs-exploitation resolved by switching algorithms rather than by tuning one.

T3, and the reason is structural: the person who modified the algorithm is the person who owns the benchmark it now leads. That’s not an accusation of bad faith — the code is published, the weaknesses are stated plainly — but a self-modified entrant topping a self-administered suite is the one result in this corpus that most needs an outside run. See synthesis for how this lands against the no-free-lunch-theorem framing.

Related

andrey-dik · population-optimization-benchmark · evolution-strategies · grey-wolf-optimizer · exploration-vs-exploitation · no-free-lunch-theorem · metaheuristic-optimization