Funk2: A Distributed Processing Language for Reflective Tracing of a Large Critic-Selector Cognitive Architecture

Bo Morgan
2010 2010 Fourth IEEE International Conference on Self-Adaptive and Self-Organizing Systems Workshop  
We see the field of metareasoning to be the answer to many large organizational problems encountered when putting together an understandable cognitive architecture, capable of commonsense reasoning. In this paper we review the EM1 implementation of the Emotion Machine critic-selector architecture, as well as explain the current progress we have made in redesigning this first version implementation. For this purpose of redesign and large-scale implementation, we have written a novel programming
more » ... anguage, Funk2, that focuses on efficient metareasoning and procedural reflection, the keystones of the critic-selector architecture. We present an argument for why the Funk2 programming language lends itself to easing the burden on programmers that prefer to not be restricted to strictly declarative programming paradigms by allowing the learning of critic and selector activation strengths by credit assignment through arbitrary procedural code. I. CLOSED-LOOP CONTROL AND LEARNING There are many artificial intelligence algorithms that provide explanations for how to accomplish goals or gather rewards in a domain. A basic artificial intelligence system consists of three processes: (1) perceptual data are generalized and categorized to learn induced abstract models, (2) abstract models are used to infer expected hypothetical states, i.e. states of future, past, or otherwise "hidden" variables, (3) actions are chosen based on considerations of different action-dependent inferences. While there are many types of machine learning algorithms that focus on this abstract 3-step closed-loop process of learning to control, the field of meta-cognition [1] focuses on making at least two layers of closed-loop systems. The first closed-loop learning algorithm learns how to deal with the external world, while the second closed-loop learning algorithm perceives the state of the algorithm below. We see meta-cognition as a layering of learning algorithms, such that the second layer algorithm learns from perceiving the activity of the first layer and controls or modifies this first layer. While it may be clear how to trace changes in the perceptual inputs of layer one of the algorithm, it is less than clear how the second layer learner should monitor the changes in the state of the first layer learner. II. A REVIEW OF THE EMOTION MACHINE V1.0 One system that implements commonsense reasoning, based on Minsky's Emotion Machine theory of mind [2], is a metareasoning system for correcting faulty plans, called EM1 (Emotion Machine, v1) [3]. EM1 is written in Lisp, using a Prolog extension as the logical resolution tool. EM1 is a layered architecture consisting of reactive, deliberative, and reflective layers. Mental critics are represented as commonsense narratives that result in queries to a collection of different Prolog knowledge bases. The commonsense narratives are given to the system in a Lisp format that is compiled into the knowledge bases as collections of horn clauses. These knowledge bases consist of collections of domain-specific horn clauses that are divided into physical, social, and mental domains of reasoning. On top of this Prolog logical substrate, the Lisp program is organized into layers as a critic-selector model of mind [4] . The narrative plans that are generated by the deliberative layer are executed by a lower-layer, called the reactive layer. Part of the reactive layer of the algorithm is written in C and runs PID control loops in a simulated social two-wheeled inverted pendulum type robot. EM1 demonstrates how a system can use commonsense narratives in order to reason by analogy in order to generate plans. Also, EM1 demonstrates a learning process that is driven by reflective critical recognition of failure. Because of the complexity of the rigid-body physics in the world, sometimes even the most carefully constructed plans fail. EM1 has a layer of reflective critics that debug deliberative narratives as they are being interpreted by using a collection of commonsense narrative debugging critics. Using narratives about social situations, EM1 infers the goals of the other agents in the world given partial knowledge of their visible physical actions. When mistakes are made in this inference process, the failure is recognized reflectively, after the fact. Specific types of debugging responses are implemented for different forms of critical failures. EM1 is a step toward a large and complex commonsense reasoning agent with multiple layers of metareasoning that inspect, control, and debug mental representations.
doi:10.1109/sasow.2010.56 dblp:conf/saso/Morgan10 fatcat:pvo3pwf46fgczakzm2pf6sx73y