1. Main Ideas of AMBR
  2. AMBR Protagonists
  3. AMBR Mechanisms
  4. Overview of a Run

AMBR Protagonists: Concepts, Instances, and Hypotheses

As any model based on the DUAL architecture, AMBR consists of nothing but agents of various kinds. They represent the knowledge and do all information processing in the model. Therefore the natural way to begin the presentation of AMBR is to introduce the various types of agents used by it.

Each AMBR agent is a DUAL agent and as such has a micro-frame. The micro-frame is a bundle of labeled slots one of which serves to designate the type of the agent. The label of this slot is type and it is filled by a list of tags such as concept, instance, hypothesis, temporary, etc. These tags are used in conjunction with one another to account for the variety of agents employed by the model. For example, the type slot of some agent can be filled by the list (temporary instance relation) thus stating that the agent in question is a temporary agent representing an instance of some relation.

There are rules that restrict the combinations among different type tags. For instance all agents of type hypothesis are also temporary. Therefore, despite the big number of possible type combinations there are only three major types of AMBR agents: concept-agent, instance-agent, and hypothesis-agent. These major types have subdivisions as illustrated in Figure 3.

agent type hierarchy

Figure 3. Main types of AMBR agents.

Concept-agents (or concepts for short) represent classes of entities. The taxonomy of classes is represented by subc and superc links between the concepts. Some concepts are classes of objects such as teapot and liquid-holder while others represent relations such as temperature-of and cause. A concept agent may also have references to some of its instances, to be associatively related (via a-link) to other concepts, etc. All concepts are permanent agents and form the backbone of AMBR's semantic memory.

Instance-agents (or instances for short) represent individual instances. Each instance agent has an inst-of slot filled by a reference to the concept agent representing the class of the instance (Figure 4). There are a several other slots with appropriate labels that relate the instance to other instances, concepts, or hypotheses. These links, like the taxonomy-oriented links mentioned above, are used by the mechanisms of the model for various purposes. Concept and instance-agents are sometimes collectivelycalled entity-agents.

      liquid-holder
        :type      (:concept
:object)
        :subc      container
        :superc    (teapot bottle
cup)
        :a-link    liquid

      teapot
        :type
(:concept :object)
        :subc      liquid-holder
        :instance
(teapot-1 tpot-73)
        :hypoth    teapot<-->bottle


teapot-1
        :type      (:instance :object)
        :subc
teapot
        :situation sit-ABC
        :hypoth    (teapot-1<-->bottle-3

teapot-1<-->bottle-4)

Figure 4. Example of concept-agents, instance-agents, and some of the links between them. Each micro-frame has additional slots (not shown in the figure). All connectionist aspects are omitted.

Concepts and instances alike are characterized by one more tag in their type list: object, relation, or situation. These tags are mutually exclusive. An object tag means that the micro-frame represents some object or a class of objects. All agents in Figure 4 belong to this category. In contrast, the relation tag designates micro-frames that represent some relation. Such micro-frames usually have S-slots that represent the arguments of the relation. The AMBR representation scheme allows to represent both specific propositions such as made-of(teapot-1, metal-1) and general propositions such as made-of(teapot,metal).

Situation-agents (or situations for short) are a special kind of instance-agents. They are distinguished by the tag situation in their type slots. Contrary to the name of the tag, such agents do not represent whole situations. Rather, they represent the spatio-temporal contiguity of a coalition of instances. Most instance agents are affiliated to some situation. The medium of this affiliation is a slot labeled situation filled by a reference to the respective situation-agent. In the example above, the agent teapot-1 is affiliated to sit-ABC. The other elements of this situation (both objects and propositions) will have the same reference in their respective slots. Thus the situation-agent that they all refer to represents the fact that all these instances have been perceived or inferred or remembered on the same occasion. On the other hand, there need not be any links from the situation agent to its elements. This is very important for the decentralized representation of situations used in AMBR. It is a whole coalition of instance-agents that represent a particular problem, scene, episode, etc. Each participant is linked to only a few other elements and no one 'knows' the situation as a whole.

The mechanisms for analogy-making try to establish correspondences between instances from different situations, between their respective concepts, and so on. These correspondences are represented in the model by correspondence-agents (not shown in Figure 3), the most important type of which are the so-called hypothesis-agents (or hypotheses for short). Each hypothesis represents a tentative correspondence between two entities based on one or more justifications. The justification of a hypothesis is the reason for its creation and maintenance by the system. In AMBR each hypothesis must have a justification. (This is one big difference between AMBR and ACME.) The justification is either semantic or structural, represented by a concept or hypothesis agent respectively.

The hypothesis-agents are organized in a constraint satisfaction network (CSN). Coherent hypotheses are connected with excitatory links while contradictory ones inhibit each other. This is the main instrument for achieving global consistency based on local computations. This approach follows the ACME model of Holyoak and Thagard (1989) but there are important differences (discussed later). Hypothesis agents have a special activation function that gives them competitive power in the CSN.

Hypothesis agents are constructed on the initiative of their justification. In the beginning of their life cycle they are created as embryo hypotheses. Those embryos that do not coincide with an existing hypothesis establish themselves and become mature hypotheses. They compete with the other hypotheses in the CSN and become either winner or loser hypotheses.

The presentation of the last few pages emphasized mostly on the symbolic declarative aspect of AMBR agents. Like all agents in the DUAL architecture, however, they are hybrid entities and have connectionist and procedural aspects as well. Different types of agents have different procedural knowledge and thus participate in the various computational mechanisms in the model.


Next page > > > AMBR Mechanisms
Previous page < < < Main Ideas of AMBR

[ DUAL Main Page ][ AMBR Main Page ] [ Back to top ]
[ Petrov's Home Page    ][ Petrov's Publications    ]
[ Kokinov's Home Page ][ Kokinov's Publications ]

Valid HTML 4.0! Page maintained by Alexander Petrov
Last updated Wednesday June 30, 1999
New Bulgarian University Center for Cognitive Science

[ Department ][ Center ][ Research ][ Programs ][ Events ][ Faculty ][ Staff ][ Students ][ Mail ]


Last updated on 2000.01.31, 08:00; please report errors to