珊瑚礁优化算法（CRO）·综合运用
（3/3）·从礁石网格到收敛曲线，一篇看清 CROm 在 MT5 实盘优化里的完整落地路径
◍ 珊瑚群淘汰与坐标索引的实现细节
这段逻辑来自一个基于珊瑚礁优化的EA框架，核心是在每一代迭代中按适应度淘汰最差的个体。代码先用对称交换把 occupiedIndices 数组首尾对调，随后根据 Fd 参数计算要清除的珊瑚数量：removeCount = MathRound(Fd * occupiedCount)，若结果 ≤0 则强制保留至少 1 个，若超过已占用数则截断为 occupiedCount，避免越界清空整个礁体。 淘汰时逐位把 occupied[pos] 置 false、reefIndices[pos] 置 -1，所有写操作前都做了 pos 在 [0, totalReefSize) 的边界校验。对外汇或贵金属实盘而言，这类群体算法对参数 Fd 极敏感，Fd 设错可能让种群过早收敛或抖动不止，属高风险调参。 GetReefCoordIndex 把二维 (row,col) 压成一维索引：row * reefCols + col，越界直接返回 -1。SortAgentsByFitness 则用冒泡按适应度降序排 indices，内层比较 reefIndices 映射后的个体适应度 a[].f，交换前双重确认 idx 在 [0, popSize) 内。 直接在 MT5 里建个继承 C_AO_CRO 的测试类，把 reefRows=10、reefCols=10、Fd=0.2 灌进去，看 removeCount 是否如预期淘汰约 20% 占用位，能快速验证这套边界保护有没有漏。
if(i < occupiedCount && (occupiedCount - class="num">1 - i) < occupiedCount) class=class="str">"cmt">// Check for out-of-bounds { class="type">int temp = occupiedIndices[i]; occupiedIndices[i] = occupiedIndices[occupiedCount - class="num">1 - i]; occupiedIndices[occupiedCount - class="num">1 - i] = temp; } } class=class="str">"cmt">// Remove the worst Fd% corals class="type">int removeCount = (class="type">int)MathRound(Fd * occupiedCount); if (removeCount <= class="num">0) removeCount = class="num">1; class=class="str">"cmt">// At least one coral if (removeCount > occupiedCount) removeCount = occupiedCount; class=class="str">"cmt">// Overflow protection for (class="type">int i = class="num">0; i < removeCount; i++) { if(i < occupiedCount) class=class="str">"cmt">// Check for out-of-bounds { class="type">int pos = occupiedIndices[i]; if(pos >= class="num">0 && pos < totalReefSize) class=class="str">"cmt">// Additional check { occupied[pos] = class="kw">false; reefIndices[pos] = -class="num">1; } } } } } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">int C_AO_CRO::GetReefCoordIndex(class="type">int row, class="type">int col) { class=class="str">"cmt">// Check for out-of-bounds if (row < class="num">0 || row >= reefRows || col < class="num">0 || col >= reefCols) class="kw">return -class="num">1; class=class="str">"cmt">// Convert a two-dimensional position to a one-dimensional index class="kw">return row * reefCols + col; } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_CRO::SortAgentsByFitness(class="type">int &indices [], class="type">int &count) { class=class="str">"cmt">// Check for an empty array if (count <= class="num">0) class="kw">return; class=class="str">"cmt">// Bubble sort by decreasing fitness for (class="type">int i = class="num">0; i < count - class="num">1; i++) { for (class="type">int j = class="num">0; j < count - i - class="num">1; j++) { if (j + class="num">1 < count) class=class="str">"cmt">// Check that j+class="num">1 is not out of bounds { class="type">int idx1 = reefIndices [indices [j]]; class="type">int idx2 = reefIndices [indices [j + class="num">1]]; if (idx1 >= class="num">0 && idx1 < popSize && idx2 >= class="num">0 && idx2 < popSize) class=class="str">"cmt">// Check indices { if (a [idx1].f < a [idx2].f) { class="type">int temp = indices [j]; indices [j] = indices [j + class="num">1]; indices [j + class="num">1] = temp; } } } } } } class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————
「珊瑚礁优化的淘汰机制重构与回测对比」
原始 CRO 在三类测试函数上跑完 10000 次迭代，总分仅 1.82096（相当于满分基准的 20.23%）；其中 Megacity 离散任务在 500 规模下结果掉到 0.09366，说明原淘汰机制把差解留得太久。 问题出在旧版只做随机替换，没有把搜索压力引向精英邻域。改法很直接：每轮选定 eliteCount 个最优珊瑚和 removeCount 个最差珊瑚，用幂参数 power=0.1 的逆幂律分布，在精英解附近生成偏离量来填洞——大多数新解贴着最优，小概率跳远，兼顾开发与探索。 改完的 CROm 参数换成 50 种群 / 20 精英 / 20 淘汰 / 0.99 交叉，同样跑三组规模。5 Hilly 结果从 0.365 升到 0.785，500 Forest 从 0.168 降到 0.163 的误差值反而更小，总分拉到 3.89459（43.27%），几乎翻倍。 不过 CROm 在种群排名表里只排第 42，头部 ANS 跨邻域搜索拿 6.134（68.15%）。它在 Megacity 离散任务上仍吃力：500 规模结果 0.107，比 ANS 的 0.231 差了一截，可视化也一般。外汇与贵金属模型若想借用这类群体优化，先记住离散约束下的退化风险很高，回测必须分函数类型拆开看。
珊瑚礁优化里被吃掉的最差解
上面那张横向对比表里，CROm（珊瑚礁优化算法）综合评分 3.895、占比 43.27%，排在第 42 位，落后于 WOAm、AEFA 这类群体智能算法，但比中央力优化 CFO 的 3.668 还是要稳一点。它的核心机制之一就是在迭代中按概率 Pd 执行捕食（Depredation），把种群里最差的一部分珊瑚个体清掉，腾位置给精英附近生成的新解。 下面这段 MT5 里的 C_AO_CRO::Depredation 函数就是干这事的原生逻辑。注意它先随机抽一个 0~1 的数，小于 Pd 才触发；接着扫一遍 occupied 数组收集所有被占用的礁格索引，若全空直接 return。 //—————————————————————————————————————————————————————————————————————————————— void C_AO_CRO::Depredation () { // Apply destruction with Pd probability if (u.RNDfromCI (0, 1) < Pd) { // Find all occupied positions and their indices int occupiedIndices []; int occupiedCount = 0; for (int i = 0; i < totalReefSize; i++) { if (occupied [i]) { ArrayResize (occupiedIndices, occupiedCount + 1); occupiedIndices [occupiedCount] = i; occupiedCount++; } } // If there are no occupied positions, exit if (occupiedCount == 0) return; // Sort the indices by fitness (from best to worst) SortAgentsByFitness (occupiedIndices, occupiedCount); // Determine the number of best solutions used to generate new ones int eliteCount = (int)MathRound (0.1 * occupiedCount); // Use top 10% if (eliteCount <= 0) eliteCount = 1; // At least one elite solution if (eliteCount > occupiedCount) eliteCount = occupiedCount; // Determine the number of worst solutions to be replaced int removeCount = (int)MathRound (Fd * occupiedCount); if (removeCount <= 0) removeCount = 1; // At least one solution is replaced if (removeCount > occupiedCount - eliteCount) removeCount = occupiedCount - eliteCount; // Do not remove elite solutions // Remove the worst solutions and replace them with new ones in the vicinity of the best ones for (int i = 0; i < removeCount; i++) { // Index of the solution to be removed (from the end of the sorted array) int removeIndex = occupiedCount - 1 - i;
| if (removeIndex < 0 | removeIndex >= occupiedCount) continue; |
|---|
int posToRemove = occupiedIndices [removeIndex]; 逐行拆一下：第 4 行用 RNDfromCI 抽均匀随机数比对 Pd，决定本代是否捕食；第 7–17 行把 occupied 里为 true 的礁格压进 occupiedIndices，顺手计 occupiedCount；第 21 行占空则退出；第 24 行按适应度给占用索引排降序；第 27–31 行算精英数，默认取占用数 10% 且至少留 1 个；第 34–38 行按 Fd 算要淘汰的最差个数，并卡死不让动精英；第 41 行起从排序尾端倒着捞最差位，posToRemove 就是待清除礁格。 在 MT5 里把 Pd 和 Fd 两个参数拉出来试，Fd 设 0.2 时每代清掉约两成最差解，种群多样性可能比 Fd=0.1 掉得快但收敛更急；外汇与贵金属行情噪声大，这类算法实盘前务必用历史 tick 回测，杠杆品种高风险，参数过拟合概率不低。
class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_CRO::Depredation() { class=class="str">"cmt">// Apply destruction with Pd probability if (u.RNDfromCI(class="num">0, class="num">1) < Pd) { class=class="str">"cmt">// Find all occupied positions and their indices class="type">int occupiedIndices []; class="type">int occupiedCount = class="num">0; for (class="type">int i = class="num">0; i < totalReefSize; i++) { if (occupied [i]) { ArrayResize(occupiedIndices, occupiedCount + class="num">1); occupiedIndices [occupiedCount] = i; occupiedCount++; } } class=class="str">"cmt">// If there are no occupied positions, exit if (occupiedCount == class="num">0) class="kw">return; class=class="str">"cmt">// Sort the indices by fitness(from best to worst) SortAgentsByFitness(occupiedIndices, occupiedCount); class=class="str">"cmt">// Determine the number of best solutions used to generate new ones class="type">int eliteCount = (class="type">int)MathRound(class="num">0.1 * occupiedCount); class=class="str">"cmt">// Use top class="num">10% if (eliteCount <= class="num">0) eliteCount = class="num">1; class=class="str">"cmt">// At least one elite solution if (eliteCount > occupiedCount) eliteCount = occupiedCount; class=class="str">"cmt">// Determine the number of worst solutions to be replaced class="type">int removeCount = (class="type">int)MathRound(Fd * occupiedCount); if (removeCount <= class="num">0) removeCount = class="num">1; class=class="str">"cmt">// At least one solution is replaced if (removeCount > occupiedCount - eliteCount) removeCount = occupiedCount - eliteCount; class=class="str">"cmt">// Do not remove elite solutions class=class="str">"cmt">// Remove the worst solutions and replace them with new ones in the vicinity of the best ones for (class="type">int i = class="num">0; i < removeCount; i++) { class=class="str">"cmt">// Index of the solution to be removed(from the end of the sorted array) class="type">int removeIndex = occupiedCount - class="num">1 - i; if (removeIndex < class="num">0 || removeIndex >= occupiedCount) class="kw">continue; class="type">int posToRemove = occupiedIndices [removeIndex];
◍ 珊瑚礁算法里按幂律撒新解
这段逻辑来自一种受珊瑚礁启发的群体优化思路：先随机挑一个待清除的位置，越界就跳过；再从精英集合里按幂律偏好选一个优解当母体。power 设成 0.1 时，偏离量 deviation 会明显偏向大步长，新解更可能在远离当前最优的邻域生成，而非只做微调。 半径取各维度范围的 0.7 倍（注释写 10% 属笔误，实际系数是 0.7），符号随机正反，最终用 MathPow(rnd, 1/power) 放大随机值。对 XAUUSD 这类跳点凶的品种，这种生成法可能比纯高斯扰动更快跳出局部劣解，但也可能在窄幅震荡时浪费评估次数。 下面把核心片段拆开看：
| if(posToRemove<0 | posToRemove>=totalReefSize) continue; // 待删位置非法就跳过 |
|---|
// Choose one of the elite solutions double power=0.1; // 幂律参数，越小长尾越重 double r=u.RNDfromCI(0,1); int eliteIdx=(int)(eliteCount); if(eliteIdx>=eliteCount) eliteIdx=eliteCount-1; // 防止越界取最后一个精英 int posBest=occupiedIndices[eliteIdx];
| if(posBest<0 | posBest>=totalReefSize) continue; |
|---|
int bestAgentIdx=reefIndices[posBest];
| if(bestAgentIdx<0 | bestAgentIdx>=popSize) continue; |
|---|
occupied[posToRemove]=false; // 腾出坑位 int newAgentIdx=reefIndices[posToRemove]; if(newAgentIdx>=0&&newAgentIdx<popSize) { for(int c=0;c<coords;c++) { double radius=0.7*(rangeMax[c]-rangeMin[c]); // 实际半径=70%区间宽 double sign=u.RNDfromCI(0,1)<0.5?-1.0:1.0; // 随机方向 double deviation=sign*radius*MathPow(u.RNDfromCI(0,1),1.0/power); // 幂律拉伸 double newValue=a[bestAgentIdx].c[c]+deviation; a[newAgentIdx].c[c]=u.SeInDiSp(newValue,rangeMin[c],rangeMax[c],rangeStep[c]); // 夹回可行域 } occupied[posToRemove]=true; reefIndices[posToRemove]=newAgentIdx; } 在 MT5 里把 power 从 0.1 调到 0.3 跑同一段，能直接对比长尾厚度对收敛速度的影响；外汇与贵金属杠杆高、回撤快，任何参数优化都先用历史数据小样本验证再上实盘。
if(posToRemove<class="num">0||posToRemove>=totalReefSize) class="kw">continue; class=class="str">"cmt">// Choose one of the elite solutions class="type">class="kw">double power=class="num">0.1; class=class="str">"cmt">// Power distribution parameter class="type">class="kw">double r=u.RNDfromCI(class="num">0,class="num">1); class="type">int eliteIdx=(class="type">int)(eliteCount); if(eliteIdx>=eliteCount) eliteIdx=eliteCount-class="num">1; class="type">int posBest=occupiedIndices[eliteIdx]; if(posBest<class="num">0||posBest>=totalReefSize) class="kw">continue; class="type">int bestAgentIdx=reefIndices[posBest]; if(bestAgentIdx<class="num">0||bestAgentIdx>=popSize) class="kw">continue; class=class="str">"cmt">// Free up space for a new solution occupied[posToRemove]=class="kw">false; class=class="str">"cmt">// Generate a new solution in the neighborhood of the selected elite solution class="type">int newAgentIdx=reefIndices[posToRemove]; class=class="str">"cmt">// Use the same agent index if(newAgentIdx>=class="num">0&&newAgentIdx<popSize) { class=class="str">"cmt">// Generation via power-law distribution around the best solution for(class="type">int c=class="num">0;c<coords;c++) { class=class="str">"cmt">// Determine the search radius(can be adapted depending on progress) class="type">class="kw">double radius=class="num">0.7*(rangeMax[c]-rangeMin[c]); class=class="str">"cmt">// class="num">10% of the range class=class="str">"cmt">// Generate a value class="kw">using a power law class="type">class="kw">double sign=u.RNDfromCI(class="num">0,class="num">1)<class="num">0.5?-class="num">1.0:class="num">1.0; class=class="str">"cmt">// Random sign class="type">class="kw">double deviation=sign*radius*MathPow(u.RNDfromCI(class="num">0,class="num">1),class="num">1.0/power); class=class="str">"cmt">// New value in the neighborhood of the best one class="type">class="kw">double newValue=a[bestAgentIdx].c[c]+deviation; class=class="str">"cmt">// Limit the value to an acceptable range a[newAgentIdx].c[c]=u.SeInDiSp(newValue,rangeMin[c],rangeMax[c],rangeStep[c]); } class=class="str">"cmt">// Populate the new solution into the reef occupied[posToRemove]=true; reefIndices[posToRemove]=newAgentIdx; }
「画得少，看得清」
CROm 把珊瑚礁的演替逻辑压成了几条硬规则：只围绕精英解撒新解，用幂参数 power=0.1 的逆幂律把大多数扰动锁在最优解附近，偶尔放一个大幅偏离去探远水。搜索半径拉到可行域 70% 后，局部利用和全局探索的配比不再靠拍脑袋。 测试台给的评分区间是 0–100，100 仅是理论上限；附带的 CalculationTestResults.mqh 能直接算出排名表，读者开 MT5 跑 Test_AO_CROm.mq5 就能复现直方图里的分布。离散函数上它掉链子，外部参数也偏多，这两点决定它不适合直接搬去调参 EA 的离散开关。 作者把广播产卵、幼体定居这类算子开源在 github 的 Population-optimization-algorithms-MQL5 库里，想写自己的群体算法，抄这几类算子比从零起步省事。外汇与贵金属市场高杠杆、高波动，任何优化器给出的参数集都只是概率优势，上线前务必用历史数据压力测试。