This final section pulls everything together and shows how the AMBR computational mechanisms can be applied to the task of analogy making.
In the present version of AMBR, the work on a problem begins with a hand-coded representation of the target situation. Some of the agents that participate in the (decentralized) description of this situation are attached to the special nodes that are sources of activation in the model. The goal element(s) are attached to the goal node; some of the other elements are attached to the input node, thus mimicking the perceptual mechanism. The input list can also include elements that do not belong to the target situation, thus modeling the external context. It is possible that target elements are presented to the system not simultaneously but incrementally, giving rise to various order effects.
Once the target elements are connected to the source nodes, the associative mechanism begins to operate. The activation spreads through the long-term memory and brings relevant conceptual and episodic information to working memory. Shortly after, the marker-passing mechanism joins in, as instance-agents emit markers upon entering the WM. The markers begin propagating the active portion of the network.
Marker intersections provoke the construction of hypothesis-agents, thus triggering the constraint-satisfaction mechanism. After consulting the secretaries, the hypotheses initiate the structure-correspondence mechanism. The secretaries register more and more hypotheses and rate their relative success.
Gradually, a number of agents enter the working memory. The activation does not spread unrestricted, however, and the intensity of memory access declines as the decay of activation prevents the nodes that are too far away from passing the threshold. Usually, two or three situations are retrieved in full and a few others only partially. These are the candidates for base analogs. In addition, the relevant concept-agents are also active and ready to guide the mapping.
The associative mechanism never stops completely because agents occasionally get in or fall out of the working memory. Moreover, the associative mechanism is responsible for controlling the speed of the symbolic aspect as well as for settling the constraint satisfaction network.
Meanwhile, the marker-passing mechanism has generated several hypotheses. In turn, they have created additional hypotheses via the structure-correspondence mechanism. The CSN has thus become fairly elaborate and winning correspondences begin to emerge. The hypotheses standing for such correspondences are promoted to winners. This makes them even more active and provides strong support for the respective entities in the main network. In this way, the base situation that best matches the target is fully and unambiguously accessed. All its elements enter working memory. The instantiation mechanism adds even more elements if such are needed to better match the target.
Sooner or later all secretaries of the target promote their winners. The mapping constructed by the model can be read from the set of winner hypotheses. (In fact, the system maintains a 'working answer' throughout the whole run. It is often unnecessary to wait for the end.) The mechanisms for transfer should have been triggered at that time. They are not yet implemented in the current version of the model, however.
It should be emphasized that everything described so far happens as a result of a dynamic emergent process. There is no central executive that controls the operation of the system. Instead, a multitude of micro-agents interact with their immediate neighbors and their local activities give rise to macroscopic phenomena that an external observer could interpret as analog access, mapping, etc.
|Previous page||< < <||AMBR Mechanisms|
|[ DUAL Main Page ]||[ Back to top ]|
|[ Petrov's Home Page ][ Petrov's Publications ]|
|[ Kokinov's Home Page ][ Kokinov's Publications ]|