Let's put the theme of this chapter in connectionist terms, because that's what it reminds me of: Let's define an activation function as the mathematical relationship between a vehicle's sensor and its motor. In all of the vehicles described to this point, the activation function has been monotonic -- an increase in the signal produces a proportional increase (or decrease) in motor speed. "Obviously, there is something very simple-minded about creatures governed by such unconditioned likes or dislikes, and we can easily see how such the-more-the-merrier behavior could lead to disaster."
Braitenberg's innovation that is explored in Vehicle 4 is the introduction of more interesting activation functions -- functions that are nonlinear, discontinuous, nonmonotonic.
The first improvement is an activation function akin to what Ballard would call a value unit -- increases in signal activity produce increases in motor speed, but only to a point -- after a maximum is reached, increases in signal activity produce decreases in motor speed. In other words, the activation function can be described nicely with a Gaussian. This produces Vehicle 4a, which can produce complicated behavior. For instance, it might navigate towards a source, but then turn away when the signal becomes too strong. Or, it could orbit a source at a fixed radius. Such complicated trajectories produce behaviours to make us "feel that their motives and tastes are much too varied and intricate to be understood by the observer." Braitenberg suggests that we would feel that this vehicle is governed by INSTINCTS.
The second improvement is to take Vehicle 4a and make its activation function discontinuous. In other words, for some levels of detected signals, the motor will stop entirely, starting up again (perhaps at maximum speed) when other signal levels are detected. "A very lifelike pattern would be: no activation up to a threshold value of the stimulus, and increasing activation beyond the threshold, starting with a fixed minimum.
(Braitenberg observes that Vehicle 4b (discontinuous activation function) define a subset of the machines consistent with Vehicle 4a, because continuous activation functions can be used to approximate discontinuous ones. This is particularly true in an environment in which friction is present. What does this remind us of??)
"Whatever their origin, thresholds in some behavior patterns make a lot of difference in the eye of the observer. These creatures, the observer would say, ponder over their DECISIONS." Their behavior seems spontaneous. "Where decisions are being made, there must be a WILL to make them. Why not?"
A final note, this chapter reminds me a little bit about the power that is achieved by introducing more complex activation functions into the hidden units of a connectionist network. In connectionism, you only get interesting things happening by adding hidden units when the activation function for such units is nonlinear. Further power is accomplished by replacing monotonic, nonlinear activation functions with nonmonotonic functions.