Nils J. Nilsson

Probabilistic logic

Because many artificial intelligence applications require the ability to reason with uncertain knowledge , it is important to seek appropriate generalizations of logic for that case. We present here a semantical generalization of logic in which the truth values of sentences are probabili~ values (between 0 and 1). Our generalization applies to any logical system for which the consistency of a finite set of sentences can be established. The method described in the present paper combines logic with probability theory in such a way that probabilistic logical entaihnent reduces to ordinary logical entailment when the probabilities of all sentences are either 0 or 1.

SEARCH STRATEGIES FOR AI PRODUCTION SYSTEMS

This chapter discusses some control strategies for artificial intelligence (AI) production systems. The overall computational efficiency of an AI production system depends upon where along the informed/uninformed spectrum the control strategy falls. The behavior of the control system as it makes rule selections can be regarded as a search process. Some examples of the ways in which the control system might search for a solution are the hill-climbing method of irrevocable rule selection, exploring a surface for a maximum, and the backtracking and graph-search regimes, search processes that permitted tentative rule selection. The chapter discusses the tentative control regimes especially with commutative production systems. Some of the search methods that developed for tentative control regimes can be adapted for use with certain types of commutative production systems using irrevocable control regimes.

Evolutionary artificial intelligence

Several challenges confront the organizer of an introductory course in artificial intelligence (AI). First, one has to decide what subject matter to include. The union of everything in all of the popular AI textbooks is much too large, and the intersection undoubtedly wont include enough of what the organizer thinks important. The second challenge is how to blend the selected topics into a coherent whole. The third involves matters such as problem sets, programming exercises, laboratory work, case studies, and collateral readings. Finally, one must decide on the main purpose of the course: is it to teach AI techniques and skills, or is it to study AIs intellectual content, perhaps presenting related topics in psychology and philosophy? In this note we concentrate on the first and second of these topics---how to present a coherent view of the core subject matter of AI.

Reasoning with Uncertain Beliefs

This chapter describes the reasoning with uncertain beliefs. There are circumstances in which it would not be appropriate for an agent to invest its beliefs with total commitment. An agent may realize that not only does it merely believe P instead of knowing P , but that it does not believe P strongly. There are many situations in which humans possess and reason with uncertain beliefs. In searching for ways to formalize the idea of beliefs having strengths, one are tempted to consider a generalization of logic in which truth values can take on values intermediate between true and false. To believe P with total commitment is to assign it the value true. To disbelieve P totally is to assign P the value false. Inventing truth values between true and false allows for various degrees of partial belief. Multivalued logics have been studied sometimes with this application in mind.

STRUCTURED OBJECT REPRESENTATIONS

The representations aggregates several related predicate calculus expressions into larger structures (sometimes called units) that are identified with important objects in the subject domain of the system. When information about one of these objects is needed by the system, the appropriate unit is accessed and all of the relevant facts about the object are retrieved at once. It uses the phrase structured objects to describe these representational schemes because of the heavy emphasis on the structure of the representation. Indeed the structure carries some of the representational and computational burden. Certain operations that might otherwise have been performed by explicit rule applications can be performed in a more automatic way by mechanisms that depend on the structure of the representation.

State and Change

This chapter discusses the actions that change world states. The notion of state is central in most conceptualizations of the physical world. A state or situation, is a snapshot of the world at a given point in time. At different points in time, the world can be in different states. A variation of the world in which there are just three blocks is considered. Each block can be somewhere on the table, or on top of exactly one other block. Each block can have at most one other block immediately on top of it. Different states of this world correspond to different configurations of blocks. The value of the notion of state is that it allows describing changing worlds. In conceptualizing a changing world, one includes states as objects in the universe of discourse, and one invent functions, and relations that depend on them.

ADVANCED PLAN-GENERATING SYSTEMS

This chapter discusses two systems that can deal with interacting goals in a more sophisticated manner than STRIPS and various hierarchical methods for plan generation. RSTRIPS is a modification of STRIPS that uses a goal regression mechanism for circumventing goal interaction problems. A typical use of this mechanism prevents RSTRIPS from applying an F -rule, F1 , that would interfere with an achieved precondition, P , needed by another F-rule, F2 , occurring later in the plan. Because F2 occurs later than F1 , it must be that F2 has some additional unachieved precondition, P ′ that led to the need to apply F1 first. Instead of applying F1 , RSTRIPS rearranges the plan by regressing F through the F -rule that achieves P . Now, the achievement of the regressed P ′ will no longer interfere with P . Some of the techniques and conventions used by RSTRIPS can best be introduced while discussing an example problem in which the goals do not happen to interact.

SEARCH STRATEGIES FOR DECOMPOSABLE PRODUCTION SYSTEMS

Publisher Summary This chapter discusses search strategies for decomposable production systems. It introduces decomposable production systems and structures called AND/OR trees, for controlling their operation. The chapter describes some heuristic strategies for searching AND/OR trees and graphs. It also describes some search techniques for graphs used in game-playing systems. The chapter defines AND/OR graphs as hyper graphs. Instead of arcs connecting pairs of nodes, there are hyperarcs connecting a parent node with a set of successor nodes. These hyperarcs are called connectors. Each k-connector is directed from parent node to a set of k successor nodes.

PRODUCTION SYSTEMS AND AI

Most artificial intelligence (AI) systems display a more or less rigid separation between the standard computational components of data, operations, and control. That is, if these systems are described at an appropriate level, one can often identify a central entity that might be called and global database that is manipulated by certain well-defined operations, all under the control of some global control strategy. The chapter discusses the importance of identifying an appropriate level of description. Near the machine-code level, any neat separation into distinct components can disappear; at the top level, the complete AI system can consist of several database/operations/control modules interacting in a complex fashion. The point is that a system consisting of separate database, operations, and control components represents an appropriate metaphorical building block for constructing lucid descriptions of AI systems.

Metaknowledge and Metareasoning

This chapter discusses the metaknowledge and metareasoning. The existence of a formal description for the process of inference is important because it allows refining characterization of belief. It is inappropriate to assume that an intelligent agent believes the logical closure of the sentences in its database. It is more appropriate to define an agents beliefs as those sentences that it can derive in a given amount of time using a given inference procedure. The conceptualization of the inference process allows defining this notion of belief formally, and as a result, one can create agents capable of reasoning in detail about the inferential abilities, and beliefs of other agents. Another important use of conceptualization and vocabulary is in introspection. An intelligent agent should be able to observe, and describe its own problem-solving processes, and it should be able to understand problem solving hints provided by others.