📘
原子轨道搜索（AOS）算法：改进与拓展（基础篇）
「原子轨道搜索在MT5里的落地形态」
原子轨道搜索（AOS）本质是把粒子在原子轨道间的跃迁逻辑搬进优化器，用来替代传统的随机游走或网格遍历。在 MetaTrader 5 的算法交易框架里，它表现为一组可封装的 C++ 类，交易者可以直接挂到 EA 的优化流程中。 这套方法在 2025 年 7 月初由社区开发者 Andrey Dik 放出实测，原文页显示浏览量 475、互动 5，属于早期验证阶段，样本量还不足以下统计结论。 对实战有意义的是：它把「轨道能级」映射成参数空间的距离度量，让优化器倾向跳过明显低效区。外汇与贵金属杠杆品种波动剧烈，任何优化器都只是缩小搜索范围，不等于过滤了过拟合风险。
◍ 从原子轨道模型到可改进的搜索算子
前一篇已经把 AOS（原子轨道搜索）的底层逻辑拆开看过：它借用原子轨道的概率分布和粒子间相互作用，在复杂解空间里做随机但受约束的寻优。核心不是穷举，而是用概率密度引导采样，让搜索倾向于高潜力区域。 这一节要做的不是复述原理，而是动它的算子。AOS 本身已经能跑，但面对多峰、高维的 MT5 参数优化任务时，探索与开发的平衡仍偏保守——实测在 20 维以上问题中后期容易陷在局部密度峰。 改进思路集中在两类特有算子：一是轨道跃迁概率的动态调整，让算法在迭代后期仍保留一定长距离跳跃可能；二是引入自适应相互作用半径，避免粒子过早聚团。目标很直接：把现有 AOS 结构补强，使其在 EA 参数寻优里更扛造。外汇与贵金属市场高波动、高风险，任何优化结果都只是概率优势，需上 MT5 用历史数据复验。
把 AOS 分子模型改成个体导向
原版 AOS 把种群当分子、搜索区域当原子，电子就是具体坐标解。层能量 BEk 取层内电子能量算术平均，BSk 键取坐标平均，位移公式 Xki[t+1] = Xki[t] + αi × (βi × LEk − γi × BSk) 里 βi、γi 两个随机因子纯属冗余，外层已有 αi 做随机扰动。 直接砍掉 BSk，让电子奔着自身历史最优而不是层平均解移动；同时删掉 βi、γi，随机数生成耗时可能压到原来的三分之一，物理含义不丢。下面这段结构体与层参数计算里，红色部分就是该删的 BSk 相关行。
MQL5 / C++
逐行拆解：S_Layer 里 <span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="keyword">class="kw">struct</span> S_Layer { <span class="keyword">class="type">int</span> pc; <span class="comment">class=class="str">"cmt">// particle counter</span> <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> <span class="keyword">class="type">class="kw">double</span> BSk; <span class="comment">class=class="str">"cmt">// connection state</span></span> <span class="keyword">class="type">class="kw">double</span> BEk; <span class="comment">class=class="str">"cmt">// binding energy</span> <span class="keyword">class="type">class="kw">double</span> LEk; <span class="comment">class=class="str">"cmt">// minimum energy</span> <span class="keyword">class="type">void</span> Init() { pc = <span class="number">class="num">0</span>; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> BSk = <span class="number">class="num">0.0</span>;</span> BEk = <span class="number">class="num">0.0</span>; LEk = <span class="number">class="num">0.0</span>; } }; <span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="comment">class=class="str">"cmt">// Calculate parameters for each layer</span> <span class="keyword">class="type">void</span> C_AO_AOS::CalcLayerParams() { <span class="keyword">class="type">class="kw">double</span> energy; <span class="comment">class=class="str">"cmt">// Handle each coordinate(atom)</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> c = <span class="number">class="num">0</span>; c < coords; c++) { atoms [c].Init(maxLayers); <span class="comment">class=class="str">"cmt">// Handle each layer</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> L = <span class="number">class="num">0</span>; L < currentLayers [c]; L++) { energy = -<span class="macro">DBL_MAX</span>; <span class="comment">class=class="str">"cmt">// Handle each electron</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> e = <span class="number">class="num">0</span>; e < popSize; e++) { <span class="keyword">if</span> (electrons [e].layerID [c] == L) { atoms [c].layers [L].pc++; atoms [c].layers [L].BEk += a [e].f; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> atoms [c].layers [L].BSk += a [e].c [c];</span> <span class="keyword">if</span> (a [e].f > energy) { energy = a [e].f; atoms [c].layers [L].LEk = a [e].c [c]; } } } <span class="comment">class=class="str">"cmt">// Calculate average values for the layer</span> <span class="keyword">if</span> (atoms [c].layers [L].pc != <span class="number">class="num">0</span>) { atoms [c].layers [L].BEk /= atoms [c].layers [L].pc; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> atoms [c].layers [L].BSk /= atoms [c].layers [L].pc;</span> } } }
double BSk 存层连接态，Init 中 BSk = 0.0 初始化；CalcLayerParams 内 atoms[c].layers[L].BSk += a[e].c[c] 累加坐标，BSk /= pc 算平均——这三处红块删掉后，层只保留 BEk（能量均值）和 LEk（层内最优坐标）。
原版电子初始化用对数正态分布，仍有非零概率命中空间任一点；改成标准差固定为 8 的正态分布，小标准差偏广域探索，大标准差利全局收敛。外汇与贵金属参数优化属高风险实验，改完请在 MT5 策略测试器以 AOSm 标识跑排名表验证收敛差异。
MQL5 / C++
<span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="keyword">class="kw">struct</span> S_Layer { <span class="keyword">class="type">int</span> pc; <span class="comment">class=class="str">"cmt">// particle counter</span> <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> <span class="keyword">class="type">class="kw">double</span> BSk; <span class="comment">class=class="str">"cmt">// connection state</span></span> <span class="keyword">class="type">class="kw">double</span> BEk; <span class="comment">class=class="str">"cmt">// binding energy</span> <span class="keyword">class="type">class="kw">double</span> LEk; <span class="comment">class=class="str">"cmt">// minimum energy</span> <span class="keyword">class="type">void</span> Init() { pc = <span class="number">class="num">0</span>; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> BSk = <span class="number">class="num">0.0</span>;</span> BEk = <span class="number">class="num">0.0</span>; LEk = <span class="number">class="num">0.0</span>; } }; <span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="comment">class=class="str">"cmt">//——————————————————————————————————————————————————————————————————————————————</span> <span class="comment">class=class="str">"cmt">// Calculate parameters for each layer</span> <span class="keyword">class="type">void</span> C_AO_AOS::CalcLayerParams() { <span class="keyword">class="type">class="kw">double</span> energy; <span class="comment">class=class="str">"cmt">// Handle each coordinate(atom)</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> c = <span class="number">class="num">0</span>; c < coords; c++) { atoms [c].Init(maxLayers); <span class="comment">class=class="str">"cmt">// Handle each layer</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> L = <span class="number">class="num">0</span>; L < currentLayers [c]; L++) { energy = -<span class="macro">DBL_MAX</span>; <span class="comment">class=class="str">"cmt">// Handle each electron</span> <span class="keyword">for</span> (<span class="keyword">class="type">int</span> e = <span class="number">class="num">0</span>; e < popSize; e++) { <span class="keyword">if</span> (electrons [e].layerID [c] == L) { atoms [c].layers [L].pc++; atoms [c].layers [L].BEk += a [e].f; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> atoms [c].layers [L].BSk += a [e].c [c];</span> <span class="keyword">if</span> (a [e].f > energy) { energy = a [e].f; atoms [c].layers [L].LEk = a [e].c [c]; } } } <span class="comment">class=class="str">"cmt">// Calculate average values for the layer</span> <span class="keyword">if</span> (atoms [c].layers [L].pc != <span class="number">class="num">0</span>) { atoms [c].layers [L].BEk /= atoms [c].layers [L].pc; <span style="background-class="type">class="kw">color:rgb(class="num">249, class="num">204, class="num">202);"> atoms [c].layers [L].BSk /= atoms [c].layers [L].pc;</span> } } }
「原子优化里的连接态与电子更新细节」
种群每次迭代后，先算所有连接维度的群体均值。下面这段把 BS 数组清零，再双层循环把每个粒子在该维度的连接值累加，最后除以 popSize 得到平均连接态，数值落在 0~1 之间，反映该维度被群体“激活”的程度。 // Calculate the general state of connections ArrayInitialize (BS, 0); for (int c = 0; c < coords; c++) { for (int e = 0; e < popSize; e++) { BS [c] += a [e].c [c]; } BS [c] /= popSize; } 逐行看：ArrayInitialize 把 BS 全置 0；外层 c 跑每个坐标（维度），内层 e 跑每个粒子，把 a[e].c[c] 累加到 BS[c]；出内层后除以 popSize 得均值。MT5 里若 popSize=50、coords=10，BS 就是长度 10 的归一化向量。 Revision 方法负责刷新最优解。先扫一遍种群，若某粒子适应度 a[i].f 大于全局最优 fB 就更新 fB 和 bestIndex；同时若大于自身历史最优 a[i].fB，就把 a[i].fB 更新并 ArrayCopy 存下坐标 cB。最后 bestIndex 有效才把全局最优坐标 cB 抄过来。 电子更新是算法核心。每个粒子每个维度先抽转移概率 φ，若 φ<PR 就直接在参数范围内随机散射；否则按层能量比较走引力更新。α 从 [-1,1] 随机取，β、γ 是 [0,1] 概率权重，分别拉向全局最优和平均连接态。外汇与贵金属参数寻优用这套逻辑时，过拟合风险高，实盘前务必用历史数据多周期验证。
MQL5 / C++
class=class="str">"cmt">// Calculate the general state of connections ArrayInitialize(BS, class="num">0); for (class="type">int c = class="num">0; c < coords; c++) { for (class="type">int e = class="num">0; e < popSize; e++) { BS [c] += a [e].c [c]; } BS [c] /= popSize; } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">// Method of revising the best solutions class="type">void C_AO_AOSm::Revision() { class="type">int bestIndex = -class="num">1; class=class="str">"cmt">// Find the best solution in the current iteration for (class="type">int i = class="num">0; i < popSize; i++) { class=class="str">"cmt">// Update the global best solution if (a [i].f > fB) { fB = a [i].f; bestIndex = i; } class=class="str">"cmt">// Update the personal best solution if (a [i].f > a [i].fB) { a [i].fB = a [i].f; ArrayCopy(a [i].cB, a [i].c, class="num">0, class="num">0, WHOLE_ARRAY); } } class=class="str">"cmt">// Update the best coordinates if a better solution is found if (bestIndex != -class="num">1) ArrayCopy(cB, a [bestIndex].c, class="num">0, class="num">0, WHOLE_ARRAY); } class=class="str">"cmt">//—————————————————————————————————————————————————————————————————————————————— class=class="str">"cmt">// Update electron positions class="type">void C_AO_AOS::UpdateElectrons() { class="type">class="kw">double α; class=class="str">"cmt">// speed coefficient class="type">class="kw">double β; class=class="str">"cmt">// best solution weight coefficient class="type">class="kw">double γ; class=class="str">"cmt">// average state weight coefficient class="type">class="kw">double φ; class=class="str">"cmt">// transition probability class="type">class="kw">double newPos; class=class="str">"cmt">// new position class="type">class="kw">double LE; class=class="str">"cmt">// best energy class="type">class="kw">double BSk; class=class="str">"cmt">// connection state class="type">int lID; class=class="str">"cmt">// layer ID class=class="str">"cmt">// Handle each particle for (class="type">int p = class="num">0; p < popSize; p++) { for (class="type">int c = class="num">0; c < coords; c++) { φ = u.RNDprobab(); if (φ < PR) { class=class="str">"cmt">// Random scatter newPos = u.RNDfromCI(rangeMin [c], rangeMax [c]); } else { lID = electrons [p].layerID [c]; α = u.RNDfromCI(-class="num">1.0, class="num">1.0); β = u.RNDprobab(); γ = u.RNDprobab(); class=class="str">"cmt">// If the current particle energy is less than the average layer energy if (a [p].f < atoms [c].layers [lID].BEk) { class=class="str">"cmt">// Moving towards the global optimum---------------------------- LE = cB [c];
常见问题
可以，把分子模型改成个体导向后，每个粒子按自身历史最优更新，先用小样本回测验证稳定性再上实盘。
连接态控制个体间信息交换强度，电子更新依赖学习因子；先固定连接态为0.5再调电子步长，避免同时动多个量。
小布可直接调用内置的 AOS 变体做参数寻优与诊断，你只需喂入品种和约束，不用手写搜索算子。
大概率是电子更新没随周期重标定，不同周期噪声尺度不同，建议按周期单独标定学习率再对比。
有可能，个体导向削弱全局探索，可保留少量连接态交叉或加随机扰动来平衡收敛与跳出。