Maes Maes90a,Maes91a proposed a distributed, non-hierarchical mechanism which takes into account both internal and external stimuli in making behavioral choices. The nodes of the network represent behavior units and are connected by special purpose links, such as inhibitory links. The overall behavior of the network is an emergent property of interactions among the nodes, and of interactions between the nodes and the environment. More specifically, activation energy flows from both external sensory readings and internal motivations to different behavioral components. Different components use the links of the network to excite and inhibit each other. After some time, the activation energy accumulates in the component that represents the ``best'' choice, which is taken as the winner, given the current situation and motivational state of an agent. This mechanism was originally used, and proven successful, in solving relatively simple problems in a traditional AI setting (i.e. blocks world), such as choosing actions in a correct sequence so as to sand a board or spray paint a block [Maes1990]. By incorporating some biological aspects into the mechanism, it is able to deal with more complex action selection problems [Maes1991a].
The main advantages of Maes' action selection mechanism are: first, the activation energy is a continuous flow which allows smooth transition from behavior to behavior; second, it is more flexible and reactive as opposed to being centrally controlled; finally, the distributed structure makes the action selection process more robust.
Some limitations have also been pointed out by Maes herself and others [Sahota1994, Blumberg1994]. For example, it is not clear how to achieve global functionality using this mechanism and careful tuning of parameters is needed. However, this is common to practically all distributed architectures and is not unique to this work. Also, since sensory inputs of each node are in the form of binary predicates, potentially useful information may be discarded. Moreover, Tyrrell Tyrrell93 has made a critical investigation of the strictly non-hierarchical and distributed computational structure used by Maes. It is believed that the underlying structure for action selection, as suggested by ethologists, is intrinsically hierarchical, rather than ``flatly'' distributed. It is argued that some of the computational deficiencies due to the non-hierarchical structure of Maes' mechanism indicate that it is not well able to deal with animal-like action selection problems [Tyrrell1993a].
|Xiaoyuan Tu||January 1996|