📐
种群优化算法：Nelder-Mead（NM），或单纯形搜索方法·进阶篇
（2/3）·当梯度法失效，1965 年的单纯形思想如何用多智能体在 MT5 里跑出实盘优化
很多交易者一遇到目标函数没解析导数就退回网格暴力扫参，既慢又容易卡在局部谷底。Nelder-Mead 用一组会变形的顶点在参数空间里自己找路，本篇把它扩成多智能体版本，继续深挖上篇未尽的工程细节。
「Nelder-Mead 代理类的成员骨架」
下面这段 C_AO_NMm 类的声明，把 Nelder-Mead 单纯形法的可调参数和内部状态一次性摊开了。coords 是坐标维度，popSize 是种群规模，simplexPoints 记录单纯形顶点数，这三个整型私有成员决定了优化空间的形状。 reflectionCoeff、expansionCoeff、contractionCoeff 三个双精度成员对应反射、扩展、收缩系数，Init 方法接收外部传入的 reflectionCoeffP / expansionCoeffP / contractionCoeffP 完成初始化。在 MT5 里把反射系数设为 1.0、扩展系数 2.0、收缩系数 0.5 是常见起点，但贵金属跨周期拟合时往往要下调扩展系数以避免过冲。 私有方法里 Sorting 负责按目标值对单纯形顶点排序，CalcCentroid 算除最差点为外的质心，Reflection / Expansion / Contraction / Flying 分别对应四种顶点更新操作。SeInDiSp、RNDfromCI、Scale 是辅助函数，处理区间随机与量纲缩放。外汇与贵金属市场高杠杆、高波动，直接套用该优化器做实盘参数搜索须先以历史数据回测验证稳定性。
MQL5 / C++
const class="type">class="kw">double reflectionCoeffP, class=class="str">"cmt">//Reflection coefficient const class="type">class="kw">double expansionCoeffP, class=class="str">"cmt">//Expansion coefficient const class="type">class="kw">double contractionCoeffP); class=class="str">"cmt">//Contraction coefficient class="kw">public: class="type">void Moving(); class="kw">public: class="type">void Revision(); class=class="str">"cmt">//---------------------------------------------------------------------------- class="kw">private: class="type">int coords; class=class="str">"cmt">//coordinates number class="kw">private: class="type">int popSize; class=class="str">"cmt">//population size class="kw">private: class="type">int simplexPoints; class=class="str">"cmt">//simplex points class="kw">private: class="type">class="kw">double reflectionCoeff; class=class="str">"cmt">//Reflection coefficient class="kw">private: class="type">class="kw">double expansionCoeff; class=class="str">"cmt">//Expansion coefficient class="kw">private: class="type">class="kw">double contractionCoeff; class=class="str">"cmt">//Contraction coefficient class="kw">private: class="type">bool revision; class="kw">private: S_Point pTemp []; class="kw">private: class="type">int ind []; class="kw">private: class="type">class="kw">double val []; class="kw">private: class="type">void Sorting(S_Point &p []); class="kw">private: class="type">void CalcCentroid(S_Simplex &s, class="type">int indW); class="kw">private: class="type">void Reflection(S_Agent &agent, class="type">int indW); class="kw">private: class="type">void Expansion(S_Agent &agent); class="kw">private: class="type">void Contraction(S_Agent &agent, class="type">int indW); class="kw">private: class="type">void Flying(S_Agent &agent, class="type">int indW); class="kw">private: class="type">class="kw">double SeInDiSp(class="type">class="kw">double In, class="type">class="kw">double InMin, class="type">class="kw">double InMax, class="type">class="kw">double Step); class="kw">private: class="type">class="kw">double RNDfromCI(class="type">class="kw">double min, class="type">class="kw">double max); class="kw">private: class="type">class="kw">double Scale(class="type">class="kw">double In, class="type">class="kw">double InMIN, class="type">class="kw">double InMAX, class="type">class="kw">double OutMIN, class="type">class="kw">double OutMAX, class="type">bool revers); }; class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Init(const class="type">int coordsP, class=class="str">"cmt">//coordinates number const class="type">int popSizeP, class=class="str">"cmt">//population size const class="type">class="kw">double reflectionCoeffP, class=class="str">"cmt">//reflection coefficient const class="type">class="kw">double expansionCoeffP, class=class="str">"cmt">//expansion coefficient
◍ Nelder-Mead 代理群的初始化与游走逻辑
这段 C_AO_NMm 类的构造函数把下山单纯形法塞进了多代理框架：simplexPoints 被定为 coords+1，也就是 N 维问题需要 N+1 个顶点。反射、扩张、收缩三个系数由外部传入，MathSrand 用微秒计数重置随机种子，避免每次回测跑出同一串伪随机序列。 Moving() 在非 revision 阶段给每个代理的单纯形顶点撒点：先用 RNDfromCI 在 [rangeMin, rangeMax] 里均匀抽，再用 SeInDiSp 按 rangeStep 吸附到离散网格。若 popSize=20、coords=5，单次 Moving 会生成 20×6×5=600 个坐标赋值，MT5 策略测试器里能直接数出这个循环次数。 Revision() 才是单纯形迭代的核心：先对每个代理的顶点按适应度排序，把倒数第二和最后一格标成 indG（好点）与 indW（坏点），CalcCentroid 除掉坏点算质心，然后统一挂 reflection 操作。注意它顺手把各代理最优顶点拎出来比 fB，若优于当前全局最优就拷进 cB——这意味着群体里任何一个单纯形冒出好解都会立刻被记录。 外汇与贵金属行情的高波动会让适应度面剧烈变形，这类随机优化在实盘大概率漂移，建议先在 MT5 用历史数据跑 popSize 从 10 到 50 的对照，观察 cB 收敛稳定性再谈参数。
MQL5 / C++
const class="type">class="kw">double contractionCoeffP) class=class="str">"cmt">//contraction coefficient { MathSrand((class="type">int)GetMicrosecondCount()); class=class="str">"cmt">// reset of the generator fB = -DBL_MAX; revision = false; coords = coordsP; popSize = popSizeP; simplexPoints = coords + class="num">1; reflectionCoeff = reflectionCoeffP; expansionCoeff = expansionCoeffP; contractionCoeff = contractionCoeffP; ArrayResize(pTemp, simplexPoints); ArrayResize(ind, simplexPoints); ArrayResize(val, simplexPoints); ArrayResize(a, popSize); for (class="type">int i = class="num">0; i < popSize; i++) { a [i].Init(coords); } ArrayResize(rangeMax, coords); ArrayResize(rangeMin, coords); ArrayResize(rangeStep, coords); ArrayResize(cB, coords); } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Moving() { if (!revision) { class="type">int cnt = class="num">0; for (class="type">int i = class="num">0; i < popSize; i++) { for (class="type">int p = class="num">0; p < simplexPoints; p++) { cnt++; for (class="type">int c = class="num">0; c < coords; c++) { a [i].s.p [p].c [c] = RNDfromCI(rangeMin [c], rangeMax [c]); a [i].s.p [p].c [c] = SeInDiSp(a [i].s.p [p].c [c], rangeMin [c], rangeMax [c], rangeStep [c]); } } } } } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Revision() { class=class="str">"cmt">//---------------------------------------------------------------------------- if (!revision) { class=class="str">"cmt">//sort agent simplex points by FF value and assign BGW for (class="type">int i = class="num">0; i < popSize; i++) { Sorting(a [i].s.p); } class=class="str">"cmt">//calculate the simplex centroid for (class="type">int i = class="num">0; i < popSize; i++) { a [i].s.indG = simplexPoints - class="num">2; a [i].s.indW = simplexPoints - class="num">1; CalcCentroid(a [i].s, a [i].s.indW); } class=class="str">"cmt">//assign the next type of operation - reflection for (class="type">int i = class="num">0; i < popSize; i++) { Reflection(a [i], a [i].s.indW); a [i].s.operation = reflection; } class=class="str">"cmt">//save the best point of the agents’ simplexes as a global solution for (class="type">int i = class="num">0; i < popSize; i++) { if (a [i].s.p [class="num">0].f > fB) { fB = a [i].s.p [class="num">0].f; ArrayCopy(cB, a [i].s.p [class="num">0].c, class="num">0, class="num">0, WHOLE_ARRAY); }
反射之后的单纯形顶点替换逻辑
这段逻辑处理的是 Nelder-Mead 单纯形优化里「反射」动作结束后的分支判定。每一个个体 a[i] 若刚做过 reflection，就把反射点适应度记到 Xr.f，再跟当前最差点 Xw、最好点 Xb 比大小，决定是替换、扩张还是收缩。 当反射点比最差点好（Xr > Xw）时，直接用反射点顶掉最差点，并打上 needUpd 标记；若反射点还胜过了最好点（Xr > Xb），就立刻跑 Expansion() 做扩张并跳过后续。反过来，反射点不如次差点（Xr <= Xg）时走 Contraction() 收缩。 如果只替换了最差点、没碰最好点，就重排顶点、重算质心并再反射一次；否则调用 Flying() 让单纯形「飞」到新位置。外汇与贵金属参数优化属高风险场景，这套逻辑仅降低陷入局部最优的概率，不保证实盘收益。 下面这段是 revision 标记下的精英保留片段：若本轮迭代发现比历史最优 fB 更优的个体，就把它拷进全局最优数组 cB。
MQL5 / C++
} revision = true; class="kw">return; } class=class="str">"cmt">//---------------------------------------------------------------------------- if (revision) { class="type">int pos = -class="num">1; for (class="type">int i = class="num">0; i < popSize; i++) { if (a [i].f > fB) pos = i; } if (pos != -class="num">1) { fB = a [pos].f; ArrayCopy(cB, a [pos].c, class="num">0, class="num">0, WHOLE_ARRAY); } }
「单纯形收缩后的坐标重算逻辑」
当代理个体在上一步被判为收缩（contraction）时，代码先把它算出的收缩点目标值 Xc.f 赋给结构体，再和当前最差点的目标值比较。若 Xc.f 大于最差点，说明收缩没带来改善，就把 Xc 塞进最差位置、重新排序、重算形心并做一次反射；否则直接调用 Flying 让该点随机飞走，避免陷在局部。 CalcCentroid 的实作是按维度累加除最差点外所有顶点的坐标，再除以 coords（维度数）得到形心。注意这里分母是 coords 而非 simplexPoints-1，若你的顶点数定义和维数不一致，形心会偏，EA 优化可能倾向早熟收敛。 Reflection 与 Expansion 都用了固定系数：Xr = Xo + reflectionCoeff*(Xo-Xw)，Xe = Xo + expansionCoeff*(Xr-Xo)，算完用 SeInDiSp 把结果夹回参数上下界与步长网格。在 MT5 里把 reflectionCoeff 从 1.0 调到 1.3 左右，黄金 1 小时框的均线周期寻优可能更快跳出平坦区，但外汇与贵金属属高风险品种，参数激进取胜概率并不保证。
MQL5 / C++
else a [i].s.p [a [i].s.indW] = a [i].s.Xr; Sorting(a [i].s.p); a [i].s.indG = simplexPoints - class="num">2; a [i].s.indW = simplexPoints - class="num">1; CalcCentroid(a [i].s, a [i].s.indW); Reflection(a [i], a [i].s.indW); class="kw">continue; } class=class="str">"cmt">//++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ class=class="str">"cmt">//if there was contraction +++++++++++++++++++++++++++++++++++++++++++ if (a [i].s.operation == contraction) { a [i].s.Xc.f = a [i].f; if (a [i].s.Xc.f > a [i].s.p [a [i].s.indW].f) { a [i].s.p [a [i].s.indW] = a [i].s.Xc; Sorting(a [i].s.p); a [i].s.indG = simplexPoints - class="num">2; a [i].s.indW = simplexPoints - class="num">1; CalcCentroid(a [i].s, a [i].s.indW); Reflection(a [i], a [i].s.indW); } else Flying(a [i], a [i].s.indW); class="kw">continue; } } } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::CalcCentroid(S_Simplex &s, class="type">int indW) { class="type">class="kw">double summ = class="num">0.0; for (class="type">int c = class="num">0; c < coords; c++) { summ = class="num">0.0; for (class="type">int p = class="num">0; p < simplexPoints; p++) { if (p != indW) summ += s.p [p].c [c]; } s.c [c] = summ / coords; } } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Reflection(S_Agent &agent, class="type">int indW) { class="type">class="kw">double Xo; class="type">class="kw">double Xr; class="type">class="kw">double Xw; for (class="type">int c = class="num">0; c < coords; c++) { Xo = agent.s.c [c]; Xw = agent.s.p [indW].c [c]; class=class="str">"cmt">//Xr = Xo + RNDfromCI(class="num">0.0, reflectionCoeff) * (Xo - Xw); Xr = Xo + reflectionCoeff * (Xo - Xw); agent.s.Xr.c [c] = SeInDiSp(Xr, rangeMin [c], rangeMax [c], rangeStep [c]); agent.c [c] = agent.s.Xr.c [c]; } agent.s.operation = reflection; } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Expansion(S_Agent &agent) { class="type">class="kw">double Xo; class="type">class="kw">double Xr; class="type">class="kw">double Xe; for (class="type">int c = class="num">0; c < coords; c++) { Xo = agent.s.c [c]; Xr = agent.s.Xr.c [c]; class=class="str">"cmt">//Xe = Xo + RNDfromCI(class="num">0.0, expansionCoeff) * (Xr - Xo); Xe = Xo + expansionCoeff * (Xr - Xo); agent.s.Xe.c [c] = SeInDiSp(Xe, rangeMin [c], rangeMax [c], rangeStep [c]); agent.c [c] = agent.s.Xe.c [c]; } agent.s.operation = expansion; } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class="type">void C_AO_NMm::Contraction(S_Agent &agent, class="type">int indW) { class="type">class="kw">double Xo; class="type">class="kw">double Xw; class="type">class="kw">double Xc;
把重复劳动交给小布
这些诊断与小布盯盘的 AIGC 已内置，打开对应品种页即可看到参数空间的单纯形收敛轨迹，你只需判断哪组权重值得上实盘。
常见问题
原文实现给每个智能体配独立单纯形，多个活体并行探索，因此可归类为群体化变体，而非经典单单纯形判定搜索。
小布盯盘内置的 AIGC 诊断可识别此类无导数优化的输出形态，但策略编译仍需在 MT5 环境完成，小布负责看盘口与收敛提示。
不会，无论维度多少，排序后'好'点始终是仅次于'最差'的第二点，质心始终排除最差顶点求平均。
外汇贵金属属高风险市场，NM 可能收敛到局部最优，建议多智能体种子分散并结合样本外验证再考虑上线。