遗传算法(GA)和遗传规划(GP)是一个有趣的研究领域。

我想知道你使用GA/GP解决的具体问题,以及如果你没有自己的库/框架,你使用了什么库/框架。

问题:

你用GA/GP解决过什么问题? 你使用了哪些库/框架?

我在寻找第一手的经验,所以请不要回答,除非你有。


当前回答

There was an competition on codechef.com (great site by the way, monthly programming competitions) where one was supposed to solve an unsolveable sudoku (one should come as close as possible with as few wrong collumns/rows/etc as possible).What I would do, was to first generate a perfect sudoku and then override the fields, that have been given. From this pretty good basis on I used genetic programming to improve my solution.I couldn't think of a deterministic approach in this case, because the sudoku was 300x300 and search would've taken too long.

其他回答

As part of my undergraduate CompSci degree, we were assigned the problem of finding optimal jvm flags for the Jikes research virtual machine. This was evaluated using the Dicappo benchmark suite which returns a time to the console. I wrote a distributed gentic alogirthm that switched these flags to improve the runtime of the benchmark suite, although it took days to run to compensate for hardware jitter affecting the results. The only problem was I didn't properly learn about the compiler theory (which was the intent of the assignment).

我本可以用现有的默认标志来播种初始种群,但有趣的是,算法发现了一个与O3优化级别非常相似的配置(但实际上在许多测试中更快)。

编辑:我还用Python写了我自己的遗传算法框架,只是使用popen命令来运行各种基准测试,尽管如果不是评估作业,我会看看pyEvolve。

2004年1月,飞利浦新显示技术公司(Philips New Display Technologies)联系了我,他们正在为有史以来第一款商业电子墨水——索尼Librie——制造电子产品。索尼Librie只在日本上市,比亚马逊Kindle和其他电子墨水在美国和欧洲上市早了好几年。

飞利浦的工程师遇到了一个大问题。在产品上市的几个月前,他们在换页面时仍然会出现重影。问题是产生静电场的200个驱动器。每个驱动器都有一个特定的电压,必须设置在0到1000mv之间。但如果你改变其中一个,就会改变一切。

因此,单独优化每个驱动器的电压是不可能的。可能的值组合的数量以数十亿计,一个特殊的相机大约需要1分钟来评估一个组合。工程师们尝试了许多标准的优化技术,但都没有达到预期的效果。

首席工程师联系了我,因为我之前已经向开源社区发布了一个遗传编程库。他问全科医生/全科医生是否会帮忙,以及我是否能参与其中。我这样做了,在大约一个月的时间里,我们一起工作,我在合成数据上编写和调整GA库,他则将其集成到他们的系统中。然后,有一个周末,他们让它和真人一起直播。

接下来的周一,我收到了他和他们的硬件设计师发来的溢美之词,说没人会相信GA发现的惊人结果。就是这样。同年晚些时候,该产品上市了。

我没有为此得到一分钱,但我有“吹嘘”的权利。他们从一开始就说他们已经超出预算了,所以我在开始工作之前就知道是什么交易。这对于气体的应用是一个很好的例子。:)

我年轻时就尝试过GA。我用Python写了一个模拟器,工作原理如下。

这些基因编码了神经网络的权重。

神经网络的输入是检测触摸的“天线”。较高的数值表示非常接近,0表示不接触。

输出是两个“轮子”。如果两个轮子都向前,这个人也向前。如果轮子方向相反,他就会转向。输出的强度决定了车轮转动的速度。

生成了一个简单的迷宫。这真的很简单,甚至很愚蠢。屏幕下方是起点,上方是球门,中间有四面墙。每面墙都有一个随机的空间,所以总是有一条路。

一开始我只是随机挑选一些人(我认为他们是bug)。只要有一个人达到了目标,或者达到了时间限制,就会计算适合度。它与当时到目标的距离成反比。

然后我把它们配对,“培育”它们来创造下一代。被选择繁殖的概率与它的适应性成正比。有时,这意味着如果一个人具有非常高的相对适应性,就会与自己反复繁殖。

I thought they would develop a "left wall hugging" behavior, but they always seemed to follow something less optimal. In every experiment, the bugs converged to a spiral pattern. They would spiral outward until they touched a wall to the right. They'd follow that, then when they got to the gap, they'd spiral down (away from the gap) and around. They would make a 270 degree turn to the left, then usually enter the gap. This would get them through a majority of the walls, and often to the goal.

我添加的一个功能是在基因中放入一个颜色矢量来跟踪个体之间的相关性。几代之后,它们的颜色都是一样的,这说明我应该有更好的繁殖策略。

我试着让他们制定更好的策略。我把神经网络复杂化了——增加了记忆和其他东西。这没有用。我总是看到同样的策略。

我尝试了各种方法,比如建立单独的基因库,在100代之后才重新组合。但没有什么能促使他们采取更好的策略。也许这是不可能的。

另一个有趣的事情是绘制适应度随时间变化的图表。有明确的模式,比如最大适合度在上升之前会下降。我从未见过一本进化论的书谈到这种可能性。

我构建了一个简单的GA,用于在音乐播放时从频谱中提取有用的模式。输出用于驱动winamp插件中的图形效果。

输入:一些FFT帧(想象一个二维浮点数组) 输出:单个浮点值(输入的加权和),阈值为0.0或1.0 基因:输入权重 适应度函数:占空比、脉宽、BPM在合理范围内的组合。

我将一些ga调整到频谱的不同部分以及不同的BPM限制,所以它们不会趋向于收敛到相同的模式。来自每个种群的前4个的输出被发送到渲染引擎。

一个有趣的副作用是,整个人群的平均健康状况是音乐变化的一个很好的指标,尽管通常需要4-5秒才能发现。

没有家庭作业。

1995年,我作为专业程序员的第一份工作是为标准普尔500指数期货编写一个基于遗传算法的自动交易系统。该应用程序是用Visual Basic 3 [!我不知道我当时是怎么做的,因为VB3甚至没有课程。

The application started with a population of randomly-generated fixed-length strings (the "gene" part), each of which corresponded to a specific shape in the minute-by-minute price data of the S&P500 futures, as well as a specific order (buy or sell) and stop-loss and stop-profit amounts. Each string (or "gene") had its profit performance evaluated by a run through 3 years of historical data; whenever the specified "shape" matched the historical data, I assumed the corresponding buy or sell order and evaluated the trade's result. I added the caveat that each gene started with a fixed amount of money and could thus potentially go broke and be removed from the gene pool entirely.

在对种群的每一次评估之后,幸存者被随机杂交(通过混合来自两个亲本的片段),一个基因被选择为亲本的可能性与它产生的利润成正比。我还添加了点突变的可能性,让事情变得有趣一点。经过几百代这样的基因,我最终得到了一个基因群,它可以把5000美元变成平均约10000美元,而且没有死亡/破碎的可能性(当然是在历史数据上)。

Unfortunately, I never got the chance to use this system live, since my boss lost close to $100,000 in less than 3 months trading the traditional way, and he lost his willingness to continue with the project. In retrospect, I think the system would have made huge profits - not because I was necessarily doing anything right, but because the population of genes that I produced happened to be biased towards buy orders (as opposed to sell orders) by about a 5:1 ratio. And as we know with our 20/20 hindsight, the market went up a bit after 1995.