Bounded Seed-AGI

Abstract

Four principal features of autonomous control systems are left both unaddressed and unaddressable by present-day engineering methodologies: (1) The ability to operate effectively in environments that are only partially known at design time; (2) A level of generality that allows a system to reassess and redefine the fulfillment of its mission in light of unexpected constraints or other unforeseen changes in the environment; (3) The ability to operate effectively in environments of significant complexity; and (4) The ability to degrade gracefullyhow it can continue striving to achieve its main goals when resources become scarce, or in light of other expected or unexpected constraining factors that impede its progress. We describe new methodological and engineering principles for addressing these shortcomings, that we have used to design a machine that becomes increasingly better at behaving in underspecified circumstances, in a goal-directed way, on the job, by modeling itself and its environment as experience accumulates. The work provides an architectural blueprint for constructing systems with high levels of operational autonomy in underspecified circumstances, starting from only a small amount of designer-specified code-a seed. Using value-driven dynamic priority scheduling to control the parallel execution of a vast number of lines of reasoning, the system accumulates increasingly useful models of its experience, resulting in recursive self-improvement that can be autonomously sustained after the machine leaves the lab, within the boundaries imposed by its designers. A prototype system named AERA has been implemented and demonstrated to learn a complex real-world task-real-time multimodal dialogue with humans-by on-line observation. Our work presents solutions to several challenges that must be solved for achieving artificial general intelligence.

Anytime Bounded Rationality

“…Models are variable defeasible knowledge: experimental evidences trigger both their construction, deletion, and the continual revision [10], of their predictive performance:…”

“…4 shows part 5 of an actual system (called S1; [10]) that observed (in real-time) human interactions of the general form "take a [color] [shape], put it there [pointing at some location], thank you" (and variations thereof) and learned how to satisfy its mission -hearing/speaking "thank you", depending on the assigned role (interviewee or interviewer). S1's seed contains (a) a drive run, (b) a model S 0 and its context {S 1 , S 2 }, (c) sensors monitoring the state of hands, objects and utterances (color (col), position (pos), attachment (att), shape (is), belonging (bel), designation (point), speech (speak)) and, (d) effectors (commands move, grab, release, speak and point).…”

“…It posits (a) a dynamic hierarchy of revisable time-aware programs (called models) (b) exploited by concurrent inferencing jobs, that are (c) continually re-scheduled by (d) a value-driven executive under (e) bounded latencies, keeping the size of all data (inputs, inferences, programs and jobs) within (f) a fixed memory budget. This model has been implemented and tested: it constitutes the control core of our auto-catalytic, endogenous, reflective architecture (AERA; [10]), demonstrated to learn (by observation) to conduct multimodal interviews of humans in real-time, bootstrapped by a minimal seed [11]. While the learning algorithm of this system has been described in prior publications [10,11], its control strategy, the ABR model, has not been published elsewhere.…”

“…This model has been implemented and tested: it constitutes the control core of our auto-catalytic, endogenous, reflective architecture (AERA; [10]), demonstrated to learn (by observation) to conduct multimodal interviews of humans in real-time, bootstrapped by a minimal seed [11]. While the learning algorithm of this system has been described in prior publications [10,11], its control strategy, the ABR model, has not been published elsewhere.…”

Coordination mechanism for integrated design of Human-Robot Interaction scenarios

“…AKIRA [32] is a framework that uses the concept of energy pools for which different processes compete to manage computing resources consumption via thread priorities within a cognitive architecture. Similarly, AERA [33] manages computing resources through job scheduling for perception and learning. However, none of these architectures include computing resource management through attention (or any other mechanism) as part of a behavior coordination mechanism for a robot, as addressed by HBBA.…”