Acceleration of evolutionary processes by learning and extended Fisher's fundamental theorem
Acceleration of evolutionary processes by learning and extended Fisher's fundamental theorem
复制标题
通过学习和扩展费舍尔基本定理加速进化过程
DOI：
10.1103/physrevresearch.4.013069
复制
发表时间：
2022
影响因子：
4.2
通讯作者：
Tetsuya J. Kobayashi
中科院分区：
文献类型：
--
作者：
So Nakashima;Tetsuya J. Kobayashi
自然选择是解释生物有机体进化过程和设计工程系统（如遗传算法）的一个普遍而有力的概念。在传统生物学中，人们认为生物体的变化完全是由随机突变引起的，而随机突变与其祖先的经历无关。然而，越来越多的证据表明，生物体通过表观遗传状态或其他信息传递方法将其信息传递给下一代。这些信息可以使后代在不依赖选择的情况下学习适应性特征，并可能在与自然选择相结合时加速进化。自然选择和个体学习的结合对于改进遗传算法或强化学习的工程应用也变得越来越重要。虽然以前有人提出通过学习加速进化过程，但没有理论基础可以支持它。为了通过学习加速进化过程，个体应该能够学习行为以优化适应性，这是群体而不是个体的特征。目前还不清楚个体学习是否以及如何提高适应性，从而加速进化。我们也缺乏一种量化加速的方法，从而使我们能够理解、验证和预测学习的影响。在这项工作中，我们证明了智能体可以通过祖先学习加速进化过程，这种学习只使用从他们的祖先传递的信息（祖先信息）。然后我们澄清，加速的发生是因为祖先信息使代理能够估计适应度梯度。最后，为了量化加速度，我们将费雪的自然选择基本定理扩展到祖先学习。扩展的费雪基本定理将加速与个体适应度的变化联系起来，从而能够定量地理解学习何时、如何以及为什么有益。
Natural selection is a general and powerful concept to explain evolutionary processes in biological organisms and to design engineering systems such as genetic algorithms. In conventional biology, it is assumed that changes in an organism occur solely from random mutations, which are independent of its ancestors' experiences. However, there is accumulating evidence that organisms transmit their information to the next generation via epigenetic states or other methods of information transfer. This information may enable descendants to learn adaptive traits without relying on selection and may accelerate evolution when combined with natural selection. The combination of natural selection and individual learning is also becoming important for engineering applications to improve genetic algorithms or reinforcement learning. While acceleration of the evolutionary process by learning has previously been suggested, no theoretical foundation is available to support it. To accelerate evolutionary processes by learning, individuals should be able to learn behaviors to optimize fitness, which is a trait of the population rather than the individual. It has not yet been clarified whether and how individual learning can improve fitness and thereby accelerate evolution. We also lack a methodology to quantify acceleration and thus enable us to understand, verify, and predict the impacts of learning. In this work, we show that agents can accelerate the evolutionary process by ancestral learning, which employs the information transmitted only from their ancestors (ancestral information). We then clarify that acceleration occurs because ancestral information enables agents to estimate the gradient of fitness. Finally, to quantify acceleration, we extend Fisher's fundamental theorem for natural selection to ancestral learning. The extended Fisher's fundamental theorem relates acceleration to the variation of individual fitness, and thus enables a quantitative understanding of when, how, and why learning is beneficial.