前言
经济学人浓咖啡马拉松计划doing！第-99天。
喝吧
Artificial selection: how evolution could design future robots
人工选择：如何利用进化设计未来机器人
What if an algorithm could mimic hundreds of millions of years of evolution?
如果有一种算法可以模拟数百万年的进化会怎样呢？
A team from Stanford University tried to do just that.
斯坦福大学的一支团队正尝试实现这种算法。
Their results, published in Nature Communications, a journal, could suggest ways to improve artificial intelligence.
他们在Nature Communications期刊发表论文，并提出了改善人工智能的建议。
The researchers created virtual environments filled with hundreds of creepy crawlies to test the approach.
为了测试这种新方法，研究学者创建了包含上百个爬虫的虚拟环境。
The “unimals” were programmed to learn tasks, such as moving a ball or maneuvering through obstacles, while continuously evolving.
经过编程的“unimal”（虚拟生物）通过不断迭代来学习任务，这些任务包括移动小球，或者穿越障碍物。
As in the animal kingdom, the fittest member would reproduce.
就像动物世界一样，适者生存。
The offspring, which overwrote the parent, would be a physical mutation involving one genetic tweak: an extra limb for stability or a different joint for flexibility, perhaps.
后代重写父代基因，这属于物理突变。比如基因微调：增加一个肢体以利于稳定性，或者增加一个关节可能会利于灵活性。
After 10 generations the most successful unimals could master tasks twice as fast as their ancestors, despite having the same baseline intelligence.
经过10个迭代之后，在相同智力基线水平下，最成功的“unimal”掌握任务的速度是祖先的两倍。
Like in nature, a robot could adapt faster if its body and artificial mind evolve in tandem.
自然情况下，如果机器人的身体和智能同时进化，它会更快适应环境。
文中所讲的斯坦福大学团队就是李飞飞团队。
（图源：科学网）
