JaxLife: An Open-Ended Agentic Simulator

TL;DR

JaxLife is an artificial life simulator in which embodied agents, parameterized by deep neural networks, must learn to survive in an expressive world containing programmable systems, and takes a step towards studying evolved behavior in more open-ended simulations.

Abstract

Human intelligence emerged through the process of natural selection and evolution on Earth. We investigate what it would take to re-create this process in silico. While past work has often focused on low-level processes (such as simulating physics or chemistry), we instead take a more targeted approach, aiming to evolve agents that can accumulate open-ended culture and technologies across generations. Towards this, we present JaxLife: an artificial life simulator in which embodied agents, parameterized by deep neural networks, must learn to survive in an expressive world containing programmable systems. First, we describe the environment and show that it can facilitate meaningful Turing-complete computation. We then analyze the evolved emergent agents' behavior, such as rudimentary communication protocols, agriculture, and tool use. Finally, we investigate how complexity scales with the amount of compute used. We believe JaxLife takes a step towards studying evolved behavior in more open-ended simulations. Our code is available at https://github.com/luchris429/JaxLife

Figures (8)

• Figure 1: JaxLife is an ALife simulation containing agents that evolve through natural selection and programmable robots that can be requisitioned as tools. The color of the lines indicate the actions performed. Agents are red while bots are grey.
• Figure 2: The agents' network architecture. Robots, agents, and the terrain each have different encoders. Entity embeddings are processed using a self-attention block, followed by cross-attention using the agent's own embedding. This, concatenated with the terrain features, is the input to an LSTM that outputs an action vector. The terrain encoder is $1\times 1$ convolution followed by a fully connected layer.
• Figure 3: Three snapshots of manually-designed useful bots. (a) shows patrolling bots, (b) shows mass terraforming and (c) shows the transport bot, which is pushing the agents.
• Figure 4: Three snapshots of the computation of Rule 110. In this figure, terrain bits are explicitly rendered.
• Figure 5: Snapshots of the simulation at various points in time. At first, the world state is effectively random. At some point, however, agents start forming a diagonal bridge, and travel along it. This coincides with a large spike in energy usage. Later on the bridge collapses and the total energy usage declines.