Full Stack Artificial Intelligence

Computer interaction + deep reinforcement learning

This work has been reduced into the following projects: “Multiparadigm learning”, “Multi Environment Learning”, “The Multi Agent Network”, “Computatrum”, “the Computatrum Family”, “computatrum.io”, and “Limboid”. (Edit: some of those projects have also been reduced into others.) Please see corresponding projects for more details.

I propose a radical next step towards advancing artificial intelligence: combine unsupervised learning GLOM-style neural network agents in a social multiagent setting with artificial selection where each agent possesses its own Linux virtual machine with restricted access to the Internet. Agents are individually unsupervised learners which interact with their own and observe other’s virtual machines in a similar way as humans do. They can also move and communicate in a 2D multiagent space. Agents are eliminated for failing to complete information tasks while the environment cycles through periods of information surplus and minus (via Internet connectivity) to promote skill performance and diversity much as natural selection does for animal life across the seasonal year. Additionally, tasks are given with decreasing preparation time. By learning to prepare for unexpected tasks, the surviving agent community specializes and learns to be generally curious without task motivation. Since the computer is a core modality of every agent, many information tasks employed for artificial selection center on computer science — especially deep learning. The agent population is eventually employed on deeper and deeper AI problems until finally it is tasked to reproduce itself — and beyond!