In recent years, many of the intelligent systems surpass human accuracy or speed in a specific narrow application or domain. However, most of these agents are only capable of a single task and incapable of generalizing or performing minimally intelligent in other tasks. For instance, agents that won a world championship against a human cannot drive a car, and agents that can play games aren’t capable of carrying out a conversation. The future of our species lies with exploration outside of our mother planet, and such interplanetary explorations will require intelligent systems that can interact and lower the workload of the crew. Herein, we propose Sigma (Σ) (Rosenbloom, Demski, & Ustun, 2016), a cognitive architecture and intelligent system that strives to combine what has been learned from four decades of independent work on symbolic cognitive architectures and probabilistic graphical models via its graphical architecture hypothesis. Sigma’s development is driven by a combination of four desiderata: grand unification, generic cognition, functional elegance, and sufficient efficiency. In particular, Sigma leverages factor graphs towards a uniform grand unification of not only traditional cognitive capabilities but also critical non-cognitive aspects, creating unique opportunities for the construction of new kinds of cognitive models that possess a Theory of Mind and are perceptual, autonomous, interactive, affective, and adaptive. Sigma’s graphical architecture has recently been extended to handle neural networks. Sigma has quite general parameter learning capabilities, in that probabilistic, neural, and reinforcement learning capabilities all emerge from a local gradient-descent-based learning mechanism operating at the code of the architecture. The aspiration for uniform grand unification, plus this unique blend of capabilities, make Sigma a well-equipped candidate to tackle the challenge of intelligent agents for deep space exploration.
In recent years, many of the satellites control systems have started to make real-time decisions without awaiting instruction. This trend of autonomy across NASA can further be accelerated with Sigma, where many of the already developed or in-development AI modules can be integrated with Sigma and be potentially used in a variety of complex systems that NASA uses including the International Space Station as well as future spacecrafts.
A recent report on the perspective of DoD on the future of AI focuses on The Deep Learning Revolution, that includes 4 Areas of Rapid Progress besides Deep Learning: reinforcement learning, graphical models, generative models and probabilistic programming languages, and hybrid architectures. Sigma is the only cognitive architecture that combines all five of these capabilities, plus deep learning.