Companion Cognitive Systems: A Step toward Human-Level AI Kenneth D. Forbus and Thomas R. Hinrichs AI Magazine 2006
Companion Cognitive Systems Our Approach Robust Reasoning and Learning Longevity and Performance Interactivity Modeling in Companions Architecture Experiment Relate / Future Work
Companion Cognitive Systems Companions will be software aide-de-camps, collaborators with their users. Robust reasoning and learning. Longevity. Interactivity.
Our Approach Robust Reasoning and Learning Longevity and Performance analogical reasoning and learning from experience. analogical processing : SME , MAC/FAC Longevity and Performance a distributed agent architecture, hosted on cluster computers “hot-swappable” Interactivity sketch understanding Relational concept maps
Modeling in Companions Situation and Domain Models. capture the current problem and relevant knowledge about it. Task and Dialogue Models. describe the shared task and where the human/computer partnership is in working on it. User Models. capture the idiosyncratic preferences, habits, and utilities of the human partner(s). Self Models. provide the Companion’s own understanding of its operations, abilities, and preferences
First-Cut Companion Architecture
Experimental Domain : Everyday Physical Reasoning (1) 68 Bennett Mechanical Comprehension test Which wheelbarrow would be easier to lift?
Everyday Physical Reasoning (2) sketching Knowledge Entry Associate (sKEA) : interface. Visual/conceptual mappings and conventions for depicting everyday objects, modeling assumptions causal models
Everyday Physical Reasoning (3) focused on learning visual/conceptual mappings the wheel/axle relationship in a wheelbarrow being a rotational connection. explored whether accumulating examples of physical principles could enable a system to solve Bennett Mechanical Comprehension test problems.
Everyday Physical Reasoning (4)
Everyday Physical Reasoning (5)
Everyday Physical Reasoning (6)
Relate Work skill learning. Conceptual understanding and learning. analogical processing, distributed agent architecture large-scale experiments on existing hardware, Representation construction
Future Work A script-based Interaction manager SEQL agent HTN planner Incorporates probabilities in its generalizations. HTN planner to handle strategies and tactics in FreeCiv, support plan recognition, and run the Executive. Decomposing our sketching software. variety of interesting questions Self-awareness Encoding Nonlinguistic multimodal communication.