Speed

The counterpart of thoughtfulness is speed. The action selection behavior can be made faster by varying the threshold as explained above. The resulting action selection is however less `thoughtful', which means that it is less goal-oriented, less situation oriented, that it takes conflicting goals less into account and that it is less biased towards ongoing plans. Nevertheless. it may sometimes be important to react fast or it may be a wasted effort to be very thoughtful (i.e., make a lot of plans and predictions).

Fortunately, the algorithm is not complex, so that it allows speed to be obtained without sacrificing too much thoughtfulness. The algorithm does however perform some sort of `search' through a network from goal modules to executable modules, so one could argue that the algorithm suffers from the same problems as traditional AI search. More specifically, that the efficiency necessarily goes down as the number of modules involved in a plan grows (the so-called `combinatorial explosion' problem). Nevertheless, it is important to take the following counterarguments into consideration:

• The search that is going on here is of a very different nature. Actually, it resembles marker passing algorithms more than the AI notion of search. The system does not construct a search tree, nor does it maintain a current hypothetical state and partial plan. In addition, it evaluates different paths in parallel, so that it does not have to start from scratch when one path does not produce a solution, but smoothly moves from one plan to another. As a result, the computation the algorithm performs is much less costly.

• The system does not `replan' completely at every timestep. The algorithm does not reinitialize the activation-levels to zero whenever an action has been taken. This implies that it may take some time to select the first action to execute, but from then on, the network is biased towards that particular situation and set of goals. This means that it will take much less time for the following actions to be selected, in particular when little has changed in the meantime with respect to the goals or current situation.

• We believe that for real autonomous agents (e.g., mobile robots) the networks will grow `larger' instead of `longer', because typically, the agent will have more tasks/goals instead of having tasks/goals that require more actions to be taken (and therefore more `planning'). Also, large subparts may exist in the network that appear to be unconnected. As a result, the efficiency of the system will not be affected so much. Even if some paths from state matchers to goal achievers would be very long, the system would still come up with an action because it does not await a convergence in the activation levels and decreases the threshold with time. The selected action might however be non optimal.

• The same simple spreading activation rules are applied to each of the modules. In addition, there are only local and fixed links among modules. This opens interesting opportunities for a massively parallel implementation, which would imply a considerable speed up.