This page outlines the six basic mechanisms used in Ambr3: spreading activation, marker passing, constraint satisfaction, structure correspondence, rating, and instantiation. The presentation is intended to give a broad and relatively self-contained overview of these mechanisms and to show how they fit together. Chapter V of (Petrov, 1998) provides a rigorous and much more detailed coverage.
As postulated by the DUAL specification each AMBR agent has a connectionist aspect, and acts as a unit in a neural network. It receives activation from the agents that interact with it, transforms this connectionist input according to its activation function, and in turn outputs activation to other agents along weighted links. Thus there is a pattern of activation over the whole population (or network) of agents. This activation originates from some special agents and then propagates the network. There is a decay factor and various thresholds that restrict the spread of activation.
This mechanism is of paramount importance in AMBR. It provides a dynamic estimate of the relevance of each individual agent. These estimates are then used by other mechanisms for various purposes. It defines the working memory of the model by bringing some agents above the threshold while keeping irrelevant ones away. This is the foundation of access subprocess in analogy-making. Spreading activation also underlies the relaxation of the constraint satisfaction network.
Activation plays another very important role in AMBR (and DUAL in general). It is the energy supply for the symbolic aspect. More active agents work faster and are more visible to other agents. Thus changes in the pattern of activation affect everything else in the model. This makes it dynamic, emergent, and context-sensitive.
Marker passing (MP) is the symbolic counterpart of the spreading activation. It has been developed within the semantic network tradition (Quillian, 1966; Fahlman, 1979; Charniak, 1983; Hendler, 1988, 1989). In its most basic form it is a tool for answering the question, "Given two nodes in the network, is there a path between them?". The idea behind the marker passing is simple: the two nodes of origin are marked, they mark their neighbors, which in turn mark their neighbors and so forth.
AMBR markers originate in instance-agents and are then passed by concept-agents 'upward' in the class hierarchy. That is, markers can go only through links labeled inst-of and subc. For example, a marker can originate from teapot-1 and then pass through teapot, liquid-holder, container, artifact, etc. Another marker starting from bottle-7 could go through bottle and meet the first one in the concept-agent liquid-holder. The latter will detect this marker intersection and create a hypothesis that teapot-1 corresponds to bottle-7. The concept node becomes the justification of the new hypothesis. In this way, the marker passing gives rise to semantically grounded hypotheses and triggers the constraint satisfaction mechanism.
The markers accumulate in the local buffers of concept-agents and provide a record of all instances of the particular class that are active at the moment. This information is then used by other mechanisms for various purposes.
The marker-passing and structure-correspondence mechanisms create hypotheses on the basis of local information only. The constraint satisfaction mechanism is responsible to achieve consistency at the level of whole coalitions. To that end, AMBR builds a constraint satisfaction network (CSN) with appropriate links between hypotheses. The pattern of activation in the CSN then gradually reaches a stable state in which a set of hypotheses emerge as winners while all others are suppressed.
In contrast with ACME (Holyoak & Thagard, 1989), the constraint satisfaction network in AMBR is tightly interconnected with the main network. This allows seamless integration with other mechanisms in the model. For example, suppose a particular hypothesis wins the competition and becomes highly active. Part of this activation spreads to the concept-agents involved in it. When the concepts become more active they process markers faster, which will tend to generate more hypotheses of the same kind. If the hypothesis is about instances, it will activate them and they in turn will support the other instances of the same coalition, etc.
Another important property of the constraint satisfaction network in AMBR is that it is built in a decentralized and incremental fashion. Individual hypotheses come one by one in the order of their creation. (Which, by the way, reflects the system's current estimates of the relevance of the elements involved.) This poses the question of how to avoid duplication of hypotheses and to establish the links needed for the relaxation algorithm. This is the responsibility of the hypotheses themselves aided by the so called secretaries.
Each instance- or concept-agent has a secretary associated with it. The secretary is not a separate agent; it is part of the entity-agent itself. The job of the secretary is to keep track of the hypotheses involving the agent in question. It records them in the hypoth slot of the agent (cf. Figure 4) and handles hypothesis registration requests.
Whenever an embryo hypothesis is born it contacts the secretaries of its two elements and requests registration. The secretaries receive these requests, consult their records, and send secretary answers to the hypothesis. There are several kinds of answers but basically they all belong to one of the following two major types. If the new hypothesis is a duplicate of an existing one, it is advised to resign in its favor. The resigning hypothesis hands over its justification to the favorite and then fizzles out. In this way many hypotheses in the CSN have several justifications even though each of them is born with only one. The links to and from justifications are excitatory and connect the CSN with the main network.
The second major type of secretary answer is establish. It is sent to hypotheses that represent some novel correspondence. When the embryo hypothesis receives such answer it becomes mature and enters the competition with other mature hypotheses. The answer contains a list of the alternative hypotheses registered at the secretary. They are the rivals of the new one and it creates symmetrical inhibitory links with them. In this way each mature hypothesis becomes incorporated in the network. When it achieves this status it starts generating its own 'child' hypotheses via the structure correspondence mechanism.
The structure correspondence (SC) mechanism generates new hypotheses on the basis of existing ones. It is also responsible for the excitatory links between coherent hypotheses. Either way, it fosters the systematicity of the mapping that emerges out of the constraint satisfaction network (Gentner, 1983).
There are several types of structure correspondence in AMBR: bottom-up, top-down, weak, etc. They are explained in detail in (Petrov, 1998). This page only conveys the general idea by providing selected examples.
Suppose there is a mature hypothesis involving two instance agents, e.g. teapot-1<->bottle-3. The bottom-up SC will create a new embryo hypothesis at the level of concepts. Namely: teapot<->bottle. If the instances are affiliated to situations, the structure correspondence mechanism will construct an embryo hypothesis about them too, e.g. sit-ABC<->sit-XYZ. These new hypotheses are likely to coincide with ones created earlier by some other agent. In these cases the secretaries of, e.g., teapot and sit-XYZ will detect the duplication and the redundant hypotheses will be forced to resign in favor of the older ones. Still, excitatory links between teapot-1<->bottle-3 and the respective concept- and situation-level hypotheses will be established. This creates the pressure that instances of the same concept and/or the same situation are mapped consistently to instances of the other concept/ situation and vice versa.
The top-down SC applies when there is a mature hypothesis involving propositions. For instance, suppose that the agent made-of-1 represents the proposition that teapot-1 is made of metal-1. Suppose further that made-of-3 states that bottle-3 is made of glass-3. Then the hypothesis made-of-1<->made-of-3 will generate the hypotheses teapot-1<->bottle-3 and metal-1<->glass-3. (It will also generate bottom-up hypotheses like made-of<->made-of, etc.)
The hypothesis teapot-1<->bottle-3, however, is probably constructed already by the marker passing mechanism (because both are liquid holders). The secretaries will then do their job and the SC-generated embryo will resign in favor of the MP-generated mature hypothesis. In the end the latter will have two justifications: semantic and structural. This gives it better competitive power in the CSN.
Another responsibility of the secretary is to rate the relative success of each hypothesis on its secretary list. It checks at regular intervals who is the current leader among the hypotheses. That is, which one has the greatest activation level. The secretary maintains ratings for each hypothesis. Ratings are numerical values indicating how long the particular hypothesis has led the competition. When a hypothesis maintains a leading status long enough, it is promoted into winner.
Thus the rating mechanism promotes current leaders into final winners. This is done by a form of competitive learning algorithm. The secretary performs rating surveys at regular intervals. Each survey detects the leader and increases its rating at the expense of the ratings of its competitors. The magnitude of the change is proportional to the margin between the activation levels of the leader and the second best hypothesis. When a particular rating reaches some critical level, the rating mechanism triggers the promotion mechanism for the respective hypothesis.
In addition to promoting winners the rating mechanism also eliminates losers. When a particular rating drops too low and the activation level of the respective hypothesis is also low, the hypothesis is sent a fizzle message that causes it to die. Non-leader hypotheses that maintain a reasonably high activation level are kept as plausible alternatives to the leader. In this way the constraint satisfaction network is trimmed of very implausible hypotheses without ruling out any possibility a priory. This adds another dimension to the dynamics of the CSN--its topology changes both by adding and removing nodes and links.
Still another function of the rating mechanism is to trigger the instantiation mechanism upon necessity.
AMBR instantiation (also called skolemization in earlier documents) is a technique for augmenting the description of some particular episode on the basis of general semantic information. This is an advanced topic that is discussed in detail in (Petrov, 1998). This page provides only an example that conveys the overall idea.
Suppose that the target situation contains a teapot and its material is explicitly represented: teapot-1 is made of metal-1. Suppose further that teapot-1 is mapped to bottle-3 belonging to some other situation. The description of the latter, however, lacks explicit proposition about the material of bottle-3. Thus there is no counterpart of the target proposition made-of(teapot-1, metal-1).
The semantic memory, however, contains a general proposition that bottles are (usually) made of glass. These general proposition is represented by an instance of the relation made-of. This instance is not affiliated to any situation and one of its arguments is a concept-agent. For example, it might be of the form made-of(bottle, prototype-glass). This proposition is handled by AMBR mechanisms in the usual way--it emits a marker, that marker intersects in the concept-agent made-of with the marker emitted by the specific proposition in the target, the marker intersection gives rise to a hypothesis, etc. Suppose that this general hypothesis wins the competition in the constraint satisfaction network (for lack of a better alternative).
The rating mechanism detects that the leading hypothesis involves a general proposition and triggers the instantiation mechanism. The latter will construct a skolem proposition that concretizes the general proposition. In the example above, the mechanism will create skolem instances of the concepts made-of and glass. No instance of bottle is needed because the recipient situation already has one as indicated by the marker from bottle-3 stored in the local buffer of bottle. The final outcome of the instantiation is that the material of bottle-3 is taken by default to be sk-glass-3, where sk-glass-3 is a skolem instance of the concept glass. This new agent affiliates to the situation containing bottle-3. It then emits a marker, which will intersect in the concept material with the marker originating from metal-1. This will create the semantically-grounded hypothesis metal-1<->sk-glass-3 which enters the competition with high chances of success as teapot-1 is already mapped to bottle-3.
| Next page | > > > | Overview of a Run |
| Previous page | < < < | AMBR Protagonists |
| [ DUAL Main Page ] | [ Back to top ] |
| [ Petrov's Home Page ][ Petrov's Publications ] | |
| [ Kokinov's Home Page ][ Kokinov's Publications ] |
Page maintained by
Alex Petrov