In order to allow the combining of evidence from different behavioral candidates and the selection of compromised actions, roboticists Rosenblatt and Payton [Rosenblatt and Payton1989] proposed an alternative action selection process. In their mechanism, all behavior nodes express preferences for each of a set of motor actions, rather than making a decision as to which is the most suitable. The final choice is a weighted sum of all the preferences. This method was later extended by Tyrrell Tyrrell92b to form what is known as free-flow hierarchies. A free-flow hierarchy implements Rosenblatt and Payton's selection process within a hierarchical action selection architecture like that of Tinbergen Tinbergen51. All nodes in the hierarchy can influence the subsequent behavior of the agent. Activities express weighted preferences for activities lower in the hierarchy. This process propagates throughout the whole hierarchy, and as a result, instead of making a decision at each layer, a decision is only made at the lowest (i.e. action) level when the most highly preferred motor action is chosen.
Simulation results [Tyrrell1993a] show that the free-flow hierarchy outperforms Maes' mechanism and several other mechanisms. However, as Blumberg Blumberg94 points out, the relatively better performance of free-flow hierarchies may be gained at the expense of high complexity and low efficiency. This is because no focus of attention is employed in a free-flow hierarchy, and preferences need to be carefully weighted. In particular, real-time solutions may not be possible since decisions can only be made after all sensory information is processed and preferences from all components are calculated. Furthermore, all hierarchical structures suffer from a lack of flexibility in the sense that connections between components, i.e. precedences, cannot be easily altered. Free-flow hierarchies are no exception.
|Xiaoyuan Tu||January 1996|