CIS 5590.004. Artificial General Intelligence

Memory and Control

1. Inference Control

Implementing a logic in a reasoning system:

Choose a data structure for the sentences (knowledge/beliefs, questions/tasks, intermediate results),
Code the grammar rules and inference rules into executable routines,
Design a mechanism (traditionally an algorithm) to select premise(s) and inference rule(s) in each step for a given task/problem.

Separating logic and control (declarative programming, production system)

An inference process (problem solving, question answering, task processing) as searching, forward and backward

Issues in control algorithms:

combinatorial/exponential explosion
creativity and flexibility

2. Alternative Solutions

Approximation algorithm: to relax the requirements for a solution.
Heuristic search: to use an evaluation function to decide the order of exploring the options.
Anytime algorithm: to terminate by external force with increasing quality of solution.
Metareasoning: to calculate the utility of computations to balance quality and cost.
Evolutionary computation: to optimize multiple solutions through resource competition.

3. Inference Control in NARS

Working environment: novel problems (insufficient knowledge), non-exhaustive exploration (insufficient resources).

Basic ideas:

Case-by-case problem-solving
Controlled concurrency

NARS design:

Task, belief, concept (concept graph)
Bag-based storage (probabilistic priority queue + hashtable)
Architecture and working cycle (unified routine)
Resource allocation policies (dynamic attention shifting)

Refernece: OpenNARS Version 1.5.7, an implementation of NAL-1 to NAL-6

Reading

Non-Axiomatic Logic: A Model of Intelligent Reasoning, 2nd Edition, Chapter 5