Goals, as typically conceived in AI planning, provide an insufficient basis for choice of action, and hence are deficient as the sole expression of an agent's objectives. Decision-theoretic utilities offer a more adequate basis, yet lack many of the computational advantages of goals. We provide a preferential semantics for goals that grounds them in decision theory and preserves the validity of some, but not all, common goal operations performed in planning. This semantic account provides a criterion for verifying the design of goal-based planning strategies, thus providing a new framework for knowledge-level analysis of planning systems. Planning to achieve goals In the predominant AI planning paradigm, planners construct plans designed to produce states satisfying particular conditions called goals. Each goal represents a partition of possible states of the world into those satisfying and those not satisfying the goal. Though planners use goals to guide their reasoning, the crude b...

The problem of planning under uncertainty has been addressed by researchers in many different fields, adopting rather different perspectives on the problem. Unfortunately, these researchers are not always aware of the relationships among these different problem formulations, often resulting in confusion and duplicated effort. Many probabilistic planning or decision making problems can be characterized as a class of Markov decision processes that allow for significant compression in representing the underlying system dynamics. It is for this class of problems that we as experts in intensional representations are advantageously positioned to contribute efficient solution methods. This paper provides a general characterization of the representational requirements for this class of problems, and we describe how to achieve computational leverage using representations that make different types of dependency information explicit. Keywords: decision-theoretic planning, action repr...

Specification of objectives constitutes a central issue in knowledge representation for planning. Decision-theoretic approaches require that representations of objectives possess a firm semantics in terms of utility functions, yet provide the flexible compositionality needed for practical preference modeling for planning systems. Modularity, or separability in specification, is the key representational feature enabling this flexibility. In the context of utility specification, modularity corresponds exactly to well-known independence concepts from multiattribute utility theory, and leads directly to approaches for composing separate preference specifications. Ultimately, we seek to use this utilitytheoretic account to justify and improve existing mechanisms for specification of preference information, and to develop new representations exhibiting tractable specification and flexible composition of preference criteria. 1 REPRESENTING UTILITY FOR PLANNING As generally conceived, the AI...

The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the o cial policies, either expressed or implied, of any other parties.

Abstract not found

This paper describes a theory called Goal-Directed Learning (gdl) that uses the principle of decision theory to choose learning tasks. The expected utility of being able to predict various features of the environment is computed and those with highest expected utility can be used as learning goals, which an agent's inductive mechanism should form theories to predict. We present a general decision-theoretic formula for the utility of learning goals, formalizing the concept that the best learning goals are those which, if learned, would maximize the agent's expected utility. The performance element of pagoda (Probabilistic Autonomous GOal-Directed Agent), an autonomous agent design presented in (desJardins 1992), is described, and a formula is given for computing the utility of learning goals in pagoda. 1 Introduction There are many differences between the models of learning that underlie most machine learning research and the way people actually learn in the world. The models used in ...

We view planning as a search through a space of plan models. A plan model consists of a partial description of a course of action (a plan) and a set of decision models that support analysis of the plan. In this framework, model building is focussed on the development of techniques to support the incremental evaluation and construction of plans. There are special problems associated with planning that make model construnction in this context much more difficult than it might be for other applications. Since many alternative structures must be generated and evaluated while planning, exhaustive search techniques for model construction are inappropriate. We seek to develop techniques for building sparse decision structures (models) that contain enough information to allow us to make search choices without swamping the evaluator with a lot of inessential data.

This paper presents a formal framework for discrete probabilistic planning so as to extend the classical Strips planning framework. Since the semantic of Strips relies on the notion of set of formulas describing a situation, uncertainty in Strips deals with uncertain sets. An axiomatic theory is described for a new species of set such that membership to these sets can be partial. Then one builds a calculus which handles uncertainty as symbolic probabilistic degrees of membership. The algebra is then plunged in classical planning and usual definitions are given along with an example. The framework holds promise in that it allows non compoundable and non comparable uncertainty (in the uncertain case, two actions may modify the same value in a non comparable manner) and gradual truth degrees. 1 Introduction Background We refer to classical planning as the Strips planning framework. That is to say the Strips assumption (i.e. all the facts that are not modified by the performance of an ...

This deliverable is the first published result of the 1st Work Package of the MUSIC project. The main objective of this deliverable is to document the results of the state-of-the-art research performed by the involved partners in the areas of selfadaptation. Additionally, this deliverable aims to document the main requirements that emerge during the development and deployment of commercial adaptive, mobile applications, targeting ubiquitous computing environments. In the MUSIC project, we are primarily concerned with developing methods and techniques for achieving adaptable software. The main target of the desired adaptation concepts, management system and mechanisms is to support both the development and run-time management of software systems that are capable of being adapted to varying contextual situations in ubiquitous computing environments. We are aiming for an adaptation system that clearly separates and externalizes the logic of adaptation mechanisms and policies from the code defining the application logic. This will enable the provision of adaptive behaviour as generic and reusable services reliving the application programmer from much of the complex tasks of encoding adaptation management and mechanisms. In order to facilitate adaptation services as indicated above, they must be based upon appropriate concepts modelling adaptability of component-based distributed applications. Furthermore, corresponding dependable adaptation mechanisms that can be applied at run-time must support design time adaptation specification elements. The adaptation services must also include policies for deciding when and how to adapt and as well as policies for undertaking the corresponding reconfiguration in a safe and timely manner. In this respect, this document studies the state of the art in these areas, and more. Additionally, as the MUSIC project aims at providing an implementation of a middleware of which the adaptation management will be a crucial part, this document also tries to collect a set of requirements that will be needed by the applications using the middleware as well as the application developers who use the provided models and tools to create them.

The problem of planning under uncertainty has been addressed by researchers in many different fields, adopting rather different perspectives on the problem. Unfortunately, these researchers are not always aware of the relationships among these different problem formulations, often resulting in confusion and duplicated effort. Many probabilisticplanning or decision making problems can be characterized as a class of Markov decision processes that allow for significant compression in representing the underlying system dynamics. It is for this class of problems that we as experts in intensional representations are advantageously positioned to contribute efficient solution methods. This paper provides a general characterization of the representational requirements for this class of problems, and we describe how to achieve computational leverage using representations that make different types of dependency information explicit.