• No results found

2.4 Learning Qualitative Models

2.4.4 INTHELEX: learning naive physics

Esposito et al. [12] describe the application of the INTHELEX learning system to the problem of learning naive physics. The goal of the system is to replicate the kind conceptual theory revision observed in small children. For example, in early childhood, the concept of ‘force’ is explained as an innate property of big or heavy objects (they are hard to move, so they have force), as the child develops the concept is revised as an acquired property of moving objects (moving objects are given force by the agent that set them in motion).

INTHELEX is an ILP engine that learns hierarchical concepts as logic programs. The system incrementally refines concepts by matching exam- ples to existing concepts and then, if required, specializing in the case of negative examples and generalzing in the case of positive examples. The system retains all examples it has been shown and ensures that learned concepts are consistent with them all.

The system was provided with a number of positive and negative ex-

amples constructed from qualitative predicates such as weight low, weight medium, and weight high (note that the qualitative values ‘low’, ‘medium’, and

‘high’ are arbitrarily chosen). The following is a typical training example: has innate f orce(s, t):- stone(s), size high(s), weight high(s)

man(m), size high(m), weight high(m) stationary(s, t), stationary(m, t),

pushes(m, s, t)

It represents the concept that ‘a large stone has innate force because it re- mains stationary after a large man pushes it’.

The system was able to learn general concepts for both innate and ac- quired force, which it is argued, are similar to those adopted by children. The following is an example learned rule:

has innate f orce(X, T ):- size high(X), weight high(X) stationary(Y, T ), stationary(X, T ),

2.4. LEARNING QUALITATIVE MODELS 51 pushes(Y, X, T )

The approach gives some insight into how child-like learning of qual- itative models could be used as a template for intelligent agent learning. However, the rules were learned from examples drawn from a very sim- ple world in which only the relevant objects exist. Learning such rules in a more complex environment would be significantly more difficult. The system also assumes that the examples are noise free; that both positive and negative examples are avaible from which to learn; and that perfect qualitative landmarks have been discovered.

Chapter 3

Q-Systems

Knowledge representation is a key component of an intelligent agent sys- tem. The representation must be sufficiently expressive to enable the agent to represent, and to reason with, the knowledge it needs to achieve its goals. It is therefore highly dependent on the types of tasks the agent is re- quired to achieve and the types of environments the agent will operate in. An agent that must operate in a wide variety of unforeseen environments requires a more expressive representation than an agent which operates in a highly constrained environment; similarly a more expressive representa- tion is required if an agent is to describe and achieve arbitrary goals rather than some pre-determined set of tasks.

The choice of representation will not only affect the agent’s operational effectiveness but also its ability to learn new knowledge. A more expres- sive representation can be more difficult to learn because it includes more types of things and describes them in greater detail. A good representa- tion balances the level of description (expressiveness) with the ability of the agent to learn (learnability). Furthermore, an overly expressive rep- resentation is not only detrimental to learning but can also lead to overly detailed models of behaviour which can make the task of planning more difficult.

This chapter describes the ‘Q-System’ representation, a novel represen- 53

tation which describes ‘systems’ of interacting objects. The representation is designed to enable an agent to operate at a human (‘common sense’) level in arbitrary everyday environments. Q-Systems are based on non- deterministic finite state machines and aim to combine an expressive ac- tion representation with qualitative process modelling. The resulting inte- grated representation can describe both action effects and qualitative be- haviour. The following sections describe the design goals, the representa- tion details, a Q-System notation, and finally a discussion of the important design decisions.

3.1

Goals

The representation has several goals that follow from the overall goals of this research. The primary goal of the representation is to enable the agent to learn models that can be used to achieve arbitrary tasks in a wide vari- ety of everyday environments. Therefore the representation must be inde- pendent of any single problem domain and expressive enough to model arbitrary unforeseen situations. A key idea is to partition the world into independent systems of interacting objects, and this must be supported by the representation.

To achieve these goals a number of desirable features are considered es- sential characteristics of a successful representation. Conversely, a number of limitations are tolerated, either due to practical reasons or to constrain the scope of the research. The essential features and tolerable limiations are related to three distinct areas of the representation: those that apply to the representation of world state, those that apply to representation of state change, and those that apply to the representation as a whole. The features and limitations applicable to the representation of world state are: • The state of the world should be represented in terms of objects and discrete properties. It is assumed that a vision system has already

3.1. GOALS 55 determined what objects exist and can observe their properties and relationship to each other.

• It should be possible to model arbitrary information about objects. This follows from the goal of learning in arbitrary environments. The representation should not be limited to a specific set of observable features derived from a single problem ddomain. The representation cannot make any assumptions about the specific features that objects may have.

• Real valued variables should be represented qualitatively. This fol- lows from the assumption that real valued variables cannot be di- rectly observed and that qualitative measurements are used in their place. It is a goal of this research to show that qualitative descrip- tions are sufficient for everyday problem domains.

• The representation should support descriptions of generalized mod- els which can be matched to specific situations. Therefore the rep- resentation must have some way of referring to generalized objects using variables.

• The representation should allow explicit representation of groups of objects. It is assumed that the vision system can distinguish interest- ing groups of objects. Groups of objects also have properties and can behave as a single object.

• The representation should support representation of partial world states. A goal of this project is to learn models in an ‘open world’. It is unrealistic to assume that world states are completely observable and therefore the representation should support partial states. The features and limitations applicable to the representation of state change are:

• The representation should represent qualitative world dynamics, i.e. how the qualitative state of the world changes under different condi- tions (especially with regard to the agent’s actions). This will allow the agent to plan by calculating future states that satisfy its goals. • Actions will be represented atomically in terms of how states change

when they are executed. This limitation is a simplifying assumption that allows the effects of actions to be the primary focus of learning rather than their execution.

• The distinction between changes resulting directly from actions and changes resulting from world dynamics should be represented ex- plicitly. To plan effectively an agent must have knowledge of how the world changes ‘by itself’ without intervention from the agent. • Observable non-qualitative changes in real valued variables will be

represented. Certain changes in the environment will not result in a change of state because real valued variables are abstracted into discrete quantity spaces. However, these changes can be important to understanding a process and, since they can often be observed, should be explicitly representable as effects of actions.

• The probability of a given state transition will be represented. A given action may result in several different outcomes; probabilities on the various outcomes will allow the agent to plan more effectively. The features and limitations applicable to the representation as a whole are:

• Isolated groups of objects interacting together as a system should be explicitly represented. This follows directly from one of the overall project goals which is to investigate the usefulness of such partition- ing.

3.2. THE Q-SYSTEM REPRESENTATION 57