| @@ -0,0 +1,244 @@ | |||||
| # Algorithm | |||||
| --- | |||||
| 天梯分数计算算法 | |||||
| 原始记录在:<https://github.com/eesast/THUAI5/discussions/86> | |||||
| 内容如下: | |||||
| ## THUAI4 | |||||
| 关于根据队式每场比赛的分数映射到天梯分数的问题: | |||||
| 队式比赛为两队对战,每队得分的区间均为 [0, 2500]。 | |||||
| 以 tanh 函数为基础进行设计。 | |||||
| 设计原则如下: | |||||
| 1. 输的扣少量天梯分,赢的得大量天梯分 | |||||
| 2. 本就有极高天梯分数的虐本就天梯分数低的,这种降维打击现象,天梯分数涨幅极小甚至不涨天梯分 | |||||
| 3. 如果在某场比赛中,两者表现差不多,即赢的比输的得分高得不多的话,那么天梯分数涨幅也不是很高 | |||||
| 4. 如果本来天梯分数很低的,赢了天梯分数很高的,那么他得到的天梯分会较高,而另一个人,天梯分数降分稍多一些 | |||||
| 5. 如果天梯分数低的赢了天梯分数高的,但是这场比赛赢得不多的话,会把两人的分数向中间靠拢 | |||||
| 6. 总体上,赢的队伍不会降天梯分;输的队伍不会加天梯分 | |||||
| 7. 其他条件相同的情况下,在本场游戏中得分越多,加的天梯分数也越高 | |||||
| 上述原则可以保证以下两个目的的达成: | |||||
| 1. 总体来看,进行的游戏场次越多,所有队伍的平均天梯分数就越高 | |||||
| 2. 经过足够多次的游戏场次,实力有一定差距的队伍的天体分数差距逐渐拉开,实力相近的队伍的天梯分数不会差别过大,各支队伍的排名趋近于收敛 | |||||
| 用 cpp 代码编写算法代码如下(`cal` 函数): | |||||
| ```cpp | |||||
| #include <iostream> | |||||
| #include <algorithm> | |||||
| #include <cmath> | |||||
| using namespace std; | |||||
| template <typename T> | |||||
| using mypair = pair<T, T>; | |||||
| // orgScore 是天梯中两队的分数;competitionScore 是这次游戏两队的得分 | |||||
| mypair<int> cal(mypair<int> orgScore, mypair<int> competitionScore) | |||||
| { | |||||
| // 调整顺序,让第一个元素成为获胜者,便于计算 | |||||
| bool reverse = false; // 记录是否需要调整 | |||||
| if (competitionScore.first < competitionScore.second) | |||||
| { | |||||
| reverse = true; | |||||
| } | |||||
| else if (competitionScore.first == competitionScore.second) | |||||
| { | |||||
| if (orgScore.first == orgScore.second) // 完全平局,不改变天梯分数 | |||||
| { | |||||
| return orgScore; | |||||
| } | |||||
| if (orgScore.first > orgScore.second) // 本次游戏平局,但一方天梯分数高,另一方天梯分数低,需要将两者向中间略微靠拢,因此天梯分数低的定为获胜者 | |||||
| { | |||||
| reverse = true; | |||||
| } | |||||
| else | |||||
| { | |||||
| reverse = false; | |||||
| } | |||||
| } | |||||
| if (reverse) // 如果需要换,换两者的顺序 | |||||
| { | |||||
| swap(competitionScore.first, competitionScore.second); | |||||
| swap(orgScore.first, orgScore.second); | |||||
| } | |||||
| // 转成浮点数 | |||||
| mypair<double> orgScoreLf; | |||||
| mypair<double> competitionScoreLf; | |||||
| orgScoreLf.first = orgScore.first; | |||||
| orgScoreLf.second = orgScore.second; | |||||
| competitionScoreLf.first = competitionScore.first; | |||||
| competitionScoreLf.second = competitionScore.second; | |||||
| mypair<int> resScore; | |||||
| const double deltaWeight = 80.0; // 差距悬殊判断参数,比赛分差超过此值就可以认定为非常悬殊了,天梯分数增量很小,防止大佬虐菜鸡的现象造成两极分化 | |||||
| double delta = (orgScoreLf.first - orgScoreLf.second) / deltaWeight; | |||||
| cout << "Tanh delta: " << tanh(delta) << endl; | |||||
| { | |||||
| const double firstnerGet = 8e-5; // 胜利者天梯得分权值 | |||||
| const double secondrGet = 5e-6; // 失败者天梯得分权值 | |||||
| double deltaScore = 100.0; // 两队竞争分差超过多少时就认为非常大 | |||||
| double correctRate = (orgScoreLf.first - orgScoreLf.second) / 100.0; // 订正的幅度,该值越小,则在势均力敌时天梯分数改变越大 | |||||
| double correct = 0.5 * (tanh((competitionScoreLf.first - competitionScoreLf.second - deltaScore) / deltaScore - correctRate) + 1.0); // 一场比赛中,在双方势均力敌时,减小天梯分数的改变量 | |||||
| resScore.first = orgScore.first + round(competitionScoreLf.first * competitionScoreLf.first * firstnerGet * (1 - tanh(delta)) * correct); // 胜者所加天梯分 | |||||
| resScore.second = orgScore.second - round((2500.0 - competitionScoreLf.second) * (2500.0 - competitionScoreLf.second) * secondrGet * (1 - tanh(delta)) * correct); // 败者所扣天梯分,2500 为得分的最大值(THUAI4 每场得分介于 0~2500 之间) | |||||
| } | |||||
| // 如果换过,再换回来 | |||||
| if (reverse) | |||||
| { | |||||
| swap(resScore.first, resScore.second); | |||||
| } | |||||
| return resScore; | |||||
| } | |||||
| ``` | |||||
| **特别注意**:此算法是在 THUAI4 的比赛直接得分封顶为 2500 分、最低不低于 0 分的前提下设计的,因此并不一定适用于 THUAI5 的情形。 | |||||
| ## THUAI5 | |||||
| 今年把得分上限这个东西去掉了。理论上今年可以得很高很高分,但是我估计大部分比赛得分在400-600左右,最高估计1000左右。算法 借 鉴 了THUAI4,算法,换了个激活函数(正态CDF),感觉分数变化相对更好了一些? | |||||
| 代码如下: | |||||
| ~~~C++ | |||||
| #include <iostream> | |||||
| #include <algorithm> | |||||
| #include <cmath> | |||||
| using namespace std; | |||||
| template <typename T> | |||||
| using mypair = pair<T, T>; | |||||
| double PHI(double x) // THUAI3: Sigmoid; THUAI4: Tanh; THUAI5: Normal Distribution CDF | |||||
| { | |||||
| //double a1 = 0.2548292592; | |||||
| //double a2 = -0.284496736; | |||||
| //double a3 = 1.421413741; | |||||
| //double a4 = -1.453152027; | |||||
| //double a5 = 1.061405429; | |||||
| //double p = 0.3275911; | |||||
| //int sign = 1; | |||||
| //if (x < 0) | |||||
| // sign = -1; | |||||
| //x = fabs(x) / sqrt(2.0); | |||||
| //double t = 1.0 / (1.0 + p * x); | |||||
| //double y = 1.0 - ((((((a5 * t + a4) * t + a3) * t) + a2) * t) + a1) * t * exp(-x * x); | |||||
| //double cdf = 0.5 * (1.0 + sign * y); | |||||
| //return (cdf - 0.5) * 2.0; // 化到[-1,1]之间 | |||||
| return erf(x / sqrt(2)); | |||||
| } | |||||
| // orgScore 是天梯中两队的分数;competitionScore 是这次游戏两队的得分 | |||||
| mypair<int> cal(mypair<int> orgScore, mypair<int> competitionScore) | |||||
| { | |||||
| // 调整顺序,让第一个元素成为获胜者,便于计算 | |||||
| bool reverse = false; // 记录是否需要调整 | |||||
| if (competitionScore.first < competitionScore.second) | |||||
| { | |||||
| reverse = true; | |||||
| } | |||||
| else if (competitionScore.first == competitionScore.second) | |||||
| { | |||||
| if (orgScore.first == orgScore.second) // 完全平局,不改变天梯分数 | |||||
| { | |||||
| return orgScore; | |||||
| } | |||||
| if (orgScore.first > orgScore.second) // 本次游戏平局,但一方天梯分数高,另一方天梯分数低,需要将两者向中间略微靠拢,因此天梯分数低的定为获胜者 | |||||
| { | |||||
| reverse = true; | |||||
| } | |||||
| else | |||||
| { | |||||
| reverse = false; | |||||
| } | |||||
| } | |||||
| if (reverse) // 如果需要换,换两者的顺序 | |||||
| { | |||||
| swap(competitionScore.first, competitionScore.second); | |||||
| swap(orgScore.first, orgScore.second); | |||||
| } | |||||
| // 转成浮点数 | |||||
| mypair<double> orgScoreLf; | |||||
| mypair<double> competitionScoreLf; | |||||
| orgScoreLf.first = orgScore.first; | |||||
| orgScoreLf.second = orgScore.second; | |||||
| competitionScoreLf.first = competitionScore.first; | |||||
| competitionScoreLf.second = competitionScore.second; | |||||
| mypair<int> resScore; | |||||
| const double deltaWeight = 90.0; // 差距悬殊判断参数,比赛分差超过此值就可以认定为非常悬殊了,天梯分数增量很小,防止大佬虐菜鸡的现象造成两极分化 | |||||
| double delta = (orgScoreLf.first - orgScoreLf.second) / deltaWeight; | |||||
| cout << "Normal CDF delta: " << PHI(delta) << endl; | |||||
| { | |||||
| const double firstnerGet = 3e-4; // 胜利者天梯得分权值 | |||||
| const double secondrGet = 1e-4; // 失败者天梯得分权值 | |||||
| double deltaScore = 100.0; // 两队竞争分差超过多少时就认为非常大 | |||||
| double correctRate = (orgScoreLf.first - orgScoreLf.second) / 100.0; // 订正的幅度,该值越小,则在势均力敌时天梯分数改变越大 | |||||
| double correct = 0.5 * (PHI((competitionScoreLf.first - competitionScoreLf.second - deltaScore) / deltaScore - correctRate) + 1.0); // 一场比赛中,在双方势均力敌时,减小天梯分数的改变量 | |||||
| resScore.first = orgScore.first + round(competitionScoreLf.first * competitionScoreLf.first * firstnerGet * (1 - PHI(delta)) * correct); // 胜者所加天梯分 | |||||
| if (competitionScoreLf.second < 1000) | |||||
| resScore.second = orgScore.second - round((1000.0 - competitionScoreLf.second) * (1000.0 - competitionScoreLf.second) * secondrGet * (1 - PHI(delta)) * correct); // 败者所扣天梯分 | |||||
| else | |||||
| resScore.second = orgScore.second; // 败者拿1000分,已经很强了,不扣分 | |||||
| } | |||||
| // 如果换过,再换回来 | |||||
| if (reverse) | |||||
| { | |||||
| swap(resScore.first, resScore.second); | |||||
| } | |||||
| return resScore; | |||||
| } | |||||
| void Print(mypair<int> score) | |||||
| { | |||||
| std::cout << " team1: " << score.first << std::endl | |||||
| << "team2: " << score.second << std::endl; | |||||
| } | |||||
| int main() | |||||
| { | |||||
| int x, y; | |||||
| std::cout << "origin score of team 1 and 2: " << std::endl; | |||||
| std::cin >> x >> y; | |||||
| auto ori = mypair<int>(x, y); | |||||
| std::cout << "game score of team 1 and 2: " << std::endl; | |||||
| std::cin >> x >> y; | |||||
| auto sco = mypair<int>(x, y); | |||||
| Print(cal(ori, sco)); | |||||
| } | |||||
| `1000 - score`(x | |||||
| `ReLU(1000 - score)`(√ | |||||
| 防止真的超过了 1000) | |||||