第6回人工生命研究会

Recently, hardware accelerators have played an important role in advancing the state-of-the-art for deep learning, enabling rapid training of neural networks and shorter research iteration cycles for their development. Much of this progress is restricted to systems that rely on gradient descent, a highly effective optimization method when we provide it with a well-defined objective function. But in areas such as artificial life, complex systems, computational biology, and even classical physics, fitness or objective functions are usually difficult to formulate if not totally non-existent. Much of the interesting behaviors we observe take place near chaotic states, where a system is constantly transitioning between order and disorder. It can be argued that intelligent life and even civilization are all complex systems operating at the edge of chaos. If we wish to study these systems, we need more efficient methods to simulate and find solutions in them.

Artificial life contains various approaches for modeling life-as-we-know-it and life-as-it-could-be, and evolutionary computation has made great progress in developing methods for evolving artificial life creatures and generating emergent structures and behaviors. Compared to gradient descent, evolution-based methods excel in escaping from local optima and finding novel solutions with great diversity. More importantly, it is often the case that complex systems are non-differentiable, rendering gradient-based methods infeasible, and even in cases where both types of methods can be used, evolution would yield better results.

However, the progress of hardware-accelerated computational methods for evolution has not kept pace with machine learning. Much of computational evolution and simulations are still conducted using CPUs, largely ignoring the recent breakthroughs in hardware accelerators such as GPUs/TPUs. Recent work started to demonstrate the effectiveness of GPUs for evolutionary algorithms and artificial life, but so far such demonstrations are tailored for specific implementations, limiting their applicability to other cases.

In this tutorial, we show our effort to enable greater access to hardware accelerators for artificial life researchers. Specifically, we describe EvoJAX, a scalable, general purpose toolkit we developed that is open source and available publicly. Building on the JAX library, our toolkit enables evolutionary algorithms to work with artificial life models running in parallel across multiple TPU/GPUs. EvoJAX achieves very high performance by implementing the algorithm, policy and task all in NumPy, which is compiled just-in-time to run on accelerators.

We will also demonstrate how one can use the EvoJAX system and extend EvoJAX for artificial life research. We showcase several extensible examples of EvoJAX for a wide range of tasks, including “Lenia” (virtual creatures evolve inside continuous cellular automata), “Ant Colony” (virtual ants form collective behavior by following pheromone), “Flocking” (a school of fish learn to move in a coherent fashion, much like Boids). Finally, we will show that EvoJAX can find solutions to most of these tasks within minutes on GPU/TPUs, compared to hours or days when using CPUs. We believe our toolkit can significantly shorten the experimental iteration cycle for researchers working with evolutionary computation.

All of our tutorial is accompanied by live running of code in the form of Python notebooks, to show how adapting EvoJAX for novel use cases is made straightforward. With Google Colab for running Python notebooks, no extra software setup is needed.

Yujin Tang is a Senior Research Engineer in Google Brain Tokyo. His research interests include neuroevolution algorithms, reinforcement learning, robotics and their applications. He obtained his B.S in Computer Science at Shanghai Jiao Tong University and M.S in Engineering Science at Waseda University.

Yingtao Tian is a Research Scientist in Google Brain Tokyo. His research interests lie in generative models and representation learning, and their applications in image generation, natural language processing, knowledge base modeling, social network modeling, bioinformatics. Recently he particularly focused on the interdisciplinary research of creativity / humanities research and machine intelligence such as using either non-differentiable evolution strategies or differentiable operations. Prior to that, he obtained his PhD in Computer Science at Stony Brook University.

Bert Chan is a Research Engineer in Google Brain Tokyo. His research interests include artificial life and evolutionary computation. He discovered a complex adaptive system called Lenia, for which he received the Outstanding Publication of 2019 award by the International Society for Artificial Life. Bert is also an external collaborator at Inria Bordeaux, and a visiting scholar at the Institute for Advanced Study, Princeton. He received his B.Sc. in Computer Science from the Chinese University of Hong Kong and M.Sc. in Cognitive Science from Lund University.

Recently, hardware accelerators have played an important role in advancing the state-of-the-art for deep learning, enabling rapid training of neural networks and shorter research iteration cycles for their development. Much of this progress is restricted to systems that rely on gradient descent, a highly effective optimization method when we provide it with a well-defined objective function. But in areas such as artificial life, complex systems, computational biology, and even classical physics, fitness or objective functions are usually difficult to formulate if not totally non-existent. Much of the interesting behaviors we observe take place near chaotic states, where a system is constantly transitioning between order and disorder. It can be argued that intelligent life and even civilization are all complex systems operating at the edge of chaos. If we wish to study these systems, we need more efficient methods to simulate and find solutions in them.

Artificial life contains various approaches for modeling life-as-we-know-it and life-as-it-could-be, and evolutionary computation has made great progress in developing methods for evolving artificial life creatures and generating emergent structures and behaviors. Compared to gradient descent, evolution-based methods excel in escaping from local optima and finding novel solutions with great diversity. More importantly, it is often the case that complex systems are non-differentiable, rendering gradient-based methods infeasible, and even in cases where both types of methods can be used, evolution would yield better results.

However, the progress of hardware-accelerated computational methods for evolution has not kept pace with machine learning. Much of computational evolution and simulations are still conducted using CPUs, largely ignoring the recent breakthroughs in hardware accelerators such as GPUs/TPUs. Recent work started to demonstrate the effectiveness of GPUs for evolutionary algorithms and artificial life, but so far such demonstrations are tailored for specific implementations, limiting their applicability to other cases.

In this tutorial, we show our effort to enable greater access to hardware accelerators for artificial life researchers. Specifically, we describe EvoJAX, a scalable, general purpose toolkit we developed that is open source and available publicly. Building on the JAX library, our toolkit enables evolutionary algorithms to work with artificial life models running in parallel across multiple TPU/GPUs. EvoJAX achieves very high performance by implementing the algorithm, policy and task all in NumPy, which is compiled just-in-time to run on accelerators.

We will also demonstrate how one can use the EvoJAX system and extend EvoJAX for artificial life research. We showcase several extensible examples of EvoJAX for a wide range of tasks, including “Lenia” (virtual creatures evolve inside continuous cellular automata), “Ant Colony” (virtual ants form collective behavior by following pheromone), “Flocking” (a school of fish learn to move in a coherent fashion, much like Boids). Finally, we will show that EvoJAX can find solutions to most of these tasks within minutes on GPU/TPUs, compared to hours or days when using CPUs. We believe our toolkit can significantly shorten the experimental iteration cycle for researchers working with evolutionary computation.

All of our tutorial is accompanied by live running of code in the form of Python notebooks, to show how adapting EvoJAX for novel use cases is made straightforward. With Google Colab for running Python notebooks, no extra software setup is needed.