谷歌发现，机器人可以变得很暴力

技术人员仍然在努力完善人工智能和自动化技术，让它们可以在复杂世界完成一两个简单的任务，例如驾驶或测谎。但我们很容易设想在未来，许多机器人可以互相合作，解决更大的问题。在最近的试验中，Alphabet Inc的DeepMind团队就开始着手搞清学习机器如何合作——以及实际情况显示的，为何经常不合作。

这个团队设计了两个简单的游戏，重复运行了它们几千次，让控制游戏的程序学习最有效的游戏策略。在某个游戏中，两个人工智能要相互竞争，收集最多的资源（绿色“苹果”），它们也可以选择使用视频下方的黄色光速射击对手。

他们的发现，使得人们对于由人工智能驱动的不久未来，产生了一些担忧。正如团队在博文中所言，苹果收集游戏中的人工智能会在资源丰富时和平共存。不过如果苹果数量不够，它们就会迅速开始射击对方来取得优势。

更令人关注的是，无论周围有多少资源，“有能力采用更复杂战术的人工智能，都会更频繁地攻击对手，也就是表现得更不具合作性。”换句话说，人工智能为了达成目标，会毫不犹豫地妨害对手——以此类推，对于阻碍的人类也可能采取类似手段。

不过在另一个游戏里，这些程序有了更多的合作理由。如果人工智能“收集资源”所得的奖励是共享的，那么它们就会设法合作——实际上也会表现得更加聪明。

这些试验的推论值得密切关注。研究人员把这些的高度理性的人工智能机器人比作“经济人”（homo economicus）——一种完全以自身利益最大化为目的的假想人类，这也是20世纪经济学中存有瑕疵却很关键的假设之一。然而现实世界里，人类不是完全理性的。因此，对于研究人员声称他们的游戏设定能帮助我们理解经济、交通等人类系统，也有人提出怀疑。

未来的机器人本身，注定要成为追求利益最大化的纯粹存在。而在人们设计能够学习和进化的人工智能时，甚至可以采用取代最狭隘的命令，如艾萨克·阿西莫夫的“机器人三定律”的行为准则。这其中当然存在着挑战，不过DeepMind竞争算法导致的不平衡行为表明，在程序员至少还需要给机器人赋予某种形式的同情心。（财富中文网）

作者：David Z. Morris

译者：严匡正

Technologists are still working to perfect A.I. and automation technologies that can accomplish one or two tasks—say, driving or fraud detection—in a very complex world. But it’s not hard to envision a future in which multiple computer agents will work together to solve even bigger problems. In a recent experiment, Alphabet Inc's DeepMind team set out to illustrate how learning machines work together—or occasionally, as it turns out, don’t.

The team ran thousands of iterations of two simple games, allowing programs controlling them to learn the most effective strategies. In one, two A.I.s competed to gather the most resources (green ‘apples’), and could also choose whether or not to sabotage their competition with the yellow zaps seen in the video below.

What they found raises some concerns about our imminent A.I.-driven future. As the team detailed in a blog post, computer agents in the apple-gathering scenario tended to coexist peacefully when there were plenty of resources. But when apples became scarce, they quickly turned to sniping one another to get an edge.

More concerning still, “agents with the capacity to implement more complex strategies try to tag the other agent more frequently, i.e. behave less cooperatively,” regardless of how many resources were around. Smart robots, in other words, have no inherent hesitation to harm one another in pursuit of their goals—or, presumably, to do the same to any human who might get in their way.

In a different scenario, though, the programs found more reason to work together. When set to ‘hunt’ for shared rewards, A.I.s would figure out how to collaborate—and were actually more likely to do so the smarter they were.

The assumptions behind these experiments deserve some scrutiny. The researchers specifically equate their hyper-rational A.I. agents with “homo economicus”—that is, the imaginary, purely gains-maximizing human that made up the flawed core of 20th century economics. But in the real world, humans aren’t all that consistently rational, casting some doubt on the researchers’ claims that their game theory scenarios could help us understand human systems like the economy and traffic.

And future robots themselves aren’t unavoidably fated to exist as pure maximizing agents. Even learning and evolving A.I.s could be designed with behavioral guidelines that would supersede their narrowest imperatives, ala Isaac Asimov’s Three Laws of Robotics. It’s not without its challenges, but the uneven behavior of DeepMind’s competitive algorithms suggests some form of empathy should be on programmers’ to-do lists.