Levels of description

First of all, we want to recall the levels of description of Dennett (1978), dis- cussed in chapter 2. Dennett distinguishes a rational level—a level that contains a “belief-system” in relation to the actor's goals, a functional level—a level with descriptions in the terms of functions of the system, and a physical level—a level that describes the physical properties of functional components.

Following up Dennett, many other cognitive scientists/psychologists have acknowledged that the mind cannot simply be described by a physical or ra- tional level alone. Anderson (1990), see table 4.1, provides a nice comparison between the levels of description assigned by various cognitive scientists. Cog- nitive scientists are not interested much in the physical (or biological) level, be- cause the details of the physical level are still unclear. Therefore, cognitive scien- tists depict the physical level as a separate layer, i.e. “the implementation layer is an approximation to the biological level” (Anderson, 1990, p. 18).

Dennett Marr Pylyshyn Rumelhart & Newell Anderson

McClelland

(1978) (1982) (1984) (1986) (1982) (1990)

Intentional Computational Semantic Knowledge Rational

Level theory Level Level Level

Functional

Level Representationand Algorithm

Algorithm Macrotheory/ Program_{Rules Symbol Level} Algorithm Functional

Architecture MicrotheoryPDP models

Register

Implementation Transfer

Level

Physical Hardware Biological _{Device Biological}

Level implementation Level

Table 4.1: Levels of description of various cognitive scientists (Anderson, 1990)2_.

Second, we should draw a distinction between the algorithm level and the implementation level, i.e. “[t]here is a distinction to be made between (a) the mental procedures and knowledge we possess that enable use to behave adap-

tively, and (b) the mechanisms that implement these procedures and know- ledge” (Anderson, 1987, p. 468)3_.

The highest level in cognition is the semantic, intentional, knowledge or ra- tional level, i.e. according to Newell, there should be another level above the symbol or algorithm level:

The Knowledge Level Hypothesis. There exists a distinct computer systems level, lying immediately above the symbol level, which is characterized by knowledge as the medium and the principle of ra- tionality[4_{] as the law of behavior. (Newell, 1982, p. 99)}

The highest level allows one to develop psychological theories. It is not di- rectly concerned with the algorithm or implementation level, or with what the mechanisms of the mind are about.

Rather, it is about constraints on the behavior of the system in order for that behavior to be optimal. If we assume that cognition is optimized, these behavioral constraints are constraints on the mech- anisms. This level of analysis is important, because it can tell us a lot about human behavior and the mechanisms of the mind. . . . [I]t leads us to formal models of the environment from which we derive behavior. Thus, its spirit is one which focuses us on what is outside the head rather than what is inside and one which demands mathe- matical precision. (Anderson, 1990, pp. 22–23)

An example of a rational analysis is known from Dennett (1987)'s intentional level that ascribes beliefs, desires and intentions to actors:

. . . rst you decide to treat the object whose behavior is to be pre- dicted as a rational agent; then you gure out what beliefs that agent ought to have, given its place in the world and its purpose. Then you gure out what desires it ought to have, on the same considerations, and nally you predict that this rational agent will act to further its goals in the light of its beliefs. A little practical reasoning from the chosen set of beliefs and desires will in many—but not all—instances yield a decision about what the agent ought to do; this is what you predict the agent will do. (p. 17)

According to Anderson (1990), the rational analysis has some advantages. In the rst place, a rational approach has the advantage of avoiding complications of the mechanistic approach.

[First o]ne has a theory that depends on the structure of an ob- servable environment and not on the unobservable structure of the mind. . . [Second, i]t offers explanation for why the mechanisms com- pute the way they do. . . [, and third, i]t offers real guidance to the- ory construction. . . [instead of relying]. . . on very weak methods to search a very large space of psychological hypotheses. (p. 30)

3_{This is similar to Pylyshyn (1984)'s distinction in Algorithm and Functional Architecture.} 4_{For a denition of the principle of rationality, see 4.3.1: cognitive principles.}

4.1. Cognitive science

In studying Multi-Agent Systems (MAS), often the rational or intentional level is considered well-suited for describing (behavioural) systems, e.g. the game theory applied in economics or the rational (logical) agent (cf. Wooldridge, 2000) in MAS. In this dissertation, we pay attention to the intentional level, be- cause we incorporate goals and interaction with the environment/social world as well. However, we also want to incorporate the functional level—algorithm and implementation levels—besides the rational or intentional level. In other words, an actor should contain (1) representations, (2) cognitive mechanisms or algo- rithms, and (3) functional mechanisms5_{, because we argue that human behaviour}

cannot solely be explained by simplied stimulus-response behaviour. A de- sign or a model in the form of a cognitive architecture is a proven methodology in unfolding the essentials of human behaviour in a more accurate way than that provided by the behaviourist approach.

The cognitive architecture is based on the fundamental view that in cognitive science an intelligent system is not completely homogeneous, i.e. it must consist of a number of functional subsystems, or modules, that cooperate to achieve intelligent information processing and behaviour. This view is supported by empirical results in cognitive psychology (Stillings et al., 1995).

This chapter primarily focuses on the classical approach of cognition. How- ever there are also two other areas—connectionism and embodied cognition— within cognitive science that deserve attention. For instance, ACT-R, as dis- cussed in this chapter, applies connectionism-like assumptions in its associative memory. In the case of embodied cognition, we adopt the ideas of the sub- sumption (architecture) of behaviours for constructing the awareness of social situations.