The above simple attention mechanism passes only the directly required information for fulfilling the animal's current intention. While this is highly efficient, it neglects some desirable sensory information necessary for better accomplishing the task at hand, or for taking compromised actions towards satisfying more than one intention. An important characteristic of animal behavior is to be able to select compromised actions from the consideration of multiple aspects of the animal's behavior and its environment [Tyrrell1993b]. In order to capture this feature in an artificial animal's behavior model, sensory information in addition to that provided about the object of focus is required.
Take for example the case where a fish intends to avoid collisions with surrounding obstacles, among which one presents the highest danger of collision (usually the closest). In most situations, the fish would have several choices of actions that it can take to maneuver around that obstacle: it can turn left, turn right, retreat, etc. Further assume that there is another nearby obstacle (either mobile or static) which is to the left of the fish and a food source to the right. With the focusser described above, sensory information about this other obstacle and the food source will be filtered out and hence cannot contribute to the decision about the choice of action. However, by considering the presence of this additional sensory information, the obvious choice of action would be to take a right turn. If the fish retreats, it cannot take advantage of eating food. If it takes a left turn the obstacle to the left would immediately become the next ``most dangerous'' obstacle and hence, in the worst case, the fish would dither back and forth.
Another example is concerned with attempting to satisfy multiple desires. Imagine that a hungry fish is chasing floating food particles. Assume the simple case of two food particles, one located straight ahead and closer to the fish, and the other a little further and to the left of the fish. Under normal conditions, the fish will focus its attention on the food that requires the least amount of effort to get--the particle in front of it. However, suppose the fish has also a desire to mate and it sees a potential mate to its left, then a better choice (a compromised choice) of action may be to turn left--such that it can eat food while approaching the potential mate. This also depends on the relative strength of the desire to mate compared to the intention to eat.
It is clear that additional information about the environment is necessary for selecting preferred strategies of action. But how accurate and complete should additional information about the environmental objects be?
At one end of the spectrum we could use a complete, accurate description (i.e., 3D positions, velocities, etc., of all relevant objects), it is possible to decide on an `optimal' action at a given moment, such as ``turn right 23.65 degrees''. However, this will involve high computational cost and hence offset the benefits of focus of attention. In addition, the artificial fish's motor system, like those of many real animals, cannot produce motions more accurate than, say, turn roughly 45 degrees. Therefore excessively accurate motor commands tend to be futile. At the other end of the spectrum we could have no additional information whatsoever, but as we have already stated, this is undesirable. So the question remains: What level of granularity should our system have in between the two extremes? Inspired by the fact that the peripheral vision of the human eye, although of poor accuracy, serves the necessary purpose. We chose to represent the additional information as qualitative recommendations of actions, henceforth referred to as motor preferences.
Instead of passing all the sensory information about the surroundings, the focusser computes and collects a set of motor preferences. The set of motor preferences corresponds to the set of motor skills of the artificial fish, namely, , , , , , .
First let us distinguish the concept of `intention' from that of `desire'. Desires are potential candidates of the intention. At any given moment, an agent can have multiple desires, but there will be only one intention which corresponds to the strongest desire. Now if a sensory stimulus relevant to a desire/intention occurs, a value is assigned to the appropriate motor preference. A positive value represents a positive recommendation for the action, and a negative value represents negative recommendation for the action. For instance, if an obstacle is close and is in front of the fish, a negative value will be assigned to . The magnitude of the value is given by a factor representing the strength of that desire. The final value of a motor preference will be a normalized sum of all the assigned values from different desires. The summation is performed first on all the positive recommendations and then on all negative recommendations. These two summations are individually normalized and then added together. For example, suppose a fish perceives three potential mates to its left and the desire to mate has strength (we will describe how the 's are calculated in Section in Chapter ). Now further suppose that the fish also detects a predator to its left and the desire to flee has strength . Then the motor preference . Therefore, in the first example described above, there are two motor preferences: ( represents the desire to avoid collision) and ( represents the desire to eat).
|Xiaoyuan Tu||January 1996|