Braitenberg, 1984, Vehicles

Ref: Braitenberg, V. (1984). Vehicles: Experiments in synthetic psychology. Cambridge, MA: MIT Press

Vehicle 1 – Getting Around

One sensor, one motor. The force exerted by the motor will be exactly proportionate to the sensor measurement, which could respond to light, temperature, etc. The example used here is temperature.

Aristotelian physics: the speed of a moving body is proportionate to the force that drives it. Friction will allow the vehicle to come to rest in a region of low temperature where the force exerted by the motor is smaller than the frictional force. Asymmetrical friction on earth will yield complicated trajectories.

Vehicle 2 – Fear and Aggression

Two sensors and two motors. There is a third vehicle, c, where each sensor is connected to both motors, but this isbasically a fancy version of vehicle 1, so we can omit it.

Vehicle 2a (COWARD) will spend more time where there is less excitatory stimulus, and may

ram the source if directly ahead. Otherwise, the vehicle will turn away from the source.

Vehicle 2b (AGGRESSOR) will act the same as 2a if the source is straight ahead. Otherwise, it will turn toward the source and hit it. It will always hit the source frontally, unless a strong perturbation makes it turn exactly away from the source.

Vehicle 3 – Love

Here, the connections are inhibitory. The vehicles race under weak stimulation and slow down under strong stimulation. They will spend more time in the vicinity of the source than away from it, coming to rest right near the source.

Vehicle 3a orients toward the source. This vehicle LOVES the source in a permanent way.

Vehicle 3b comes to rest facing away from the source, and may not stay there long, since it now has weak stimulation. This vehicle is an EXPLORER, liking to be nearby the source, but keeping an eye out for other, perhaps stronger stimuli.

Vehicle 3c has 4 pairs of sensors, each for different environmental stimuli – e.g., light, temperature, oxygen concentration, amount of organic matter. This can create very interesting behavior. This vehicle seems to have a system of VALUES, and perhaps even KNOWLEDGE.

Vehicle 4 – Values and Special Tastes

Vehicles 1-3 used only monotonic relationships. Here we should make these relationships more complicated.

Vehicle 4a uses an upside-down U-shaped curve for its connections. Up to a point, activation of the sensor increases motor activity, but once the motor reaches a threshold speed, further activation starts to decrease the speed. Some connections can have nonlinear functions, and others can be monotonic. This will lead to complex behaviors, some of which will seem to be INSTINCTS.

Vehicle 4b will have relationships that are not smooth curves, but have definite breaks. There may be thresholds, step functions, etc. A lifelike function: no activation up to a threshold, increasing activation beyond the threshold, starting with a fixed minimum. If we have friction, thresholds will have occurred naturally. This vehicle seems to make DECISIONS.

Vehicle 5 – Logic

The Law of Uphill Analysis and Downhill Invention: It is fun and easy to create machines that do tricks, and to observe the repertoire of these machines’ behaviors. However, it is much more difficult to start from the outside and try to guess the internal structure just from the observation of behavior. Thus, analysis is more difficult than invention in the sense that induction takes more time to perform than deduction.

A psychological consequence: We when analyze a mechanism, we tend to overestimate its complexity. In uphill analysis, a given degree of complexity offers more resistance to our minds than it would if we encountered it in the downhill process of invention.

Threshold devices. To be interposed between sensors and motors or connected to each other in complexes. Two types of simple threshold devices are all-or-none. One gives no output until a threshold is reached, then full output. The other gives full output until a threshold is reached, then no output. Each can have a knob to set the threshold. Connections to threshold units can be either excitatory or inhibitory. These are basically McCullough-Pitts neurons.

This will enable this vehicle to have very specific responses to very precise stimuli. It will appear to have NAMES. It will be able to count. It will be able to solve problems in LOGIC.

This vehicle can have elementary MEMORY. After a threshold has been reached, a threshold device can activate another device that connects back with feedback, and thus keeps something happening forever. There is a limited number of facts that can be stored this way. Still, since it doesn’t have to hold all the information at a time, it can do calculations that involve bigger quantities than it can store.

Vehicle 6 – Selection, the Impersonal Engineer

Evolutionary algorithms. Have an environment of stimuli on the table. Make new vehicles by copying existing ones on the table, ignoring ones on the floor. Produce new vehicles at roughly the rate that old ones fall off to keep the population stable. Occasional mistakes will be made (mutations). We can “accidentally” combine one part of the brain from one vehicle, and another part of the brain from another vehicle. This may lead to vehicles with complicated wiring that is hard to analyze and thus feels mysterious.

Vehicle 7 – Concepts

Mnemotrix – a wire with an interesting property. Its resistance is at first very high, and stays high unless the two components that it connects are simultaneously traversed by an electric current. When this happens, the resistance of the mnemotrix decreases and remains low for a while, little by little returning to its initial value.

Now we put a piece of Mnemotrix between any two threshold devices in a complicated vehicle of type 5. This will allow it to learn by ASSOCIATION. It may even seem to form CONCEPTS, and GENERALIZATIONS.

Vehicle 8 – Space, Things, and Movement

This takes us into higher order perception. Put an array of photocells in the front of the vehicle, with a lens in front to make an eye (e.g., 10 by 10). If we connect batches of four cells in a square to a 4-threshold device, we have an object detect. If we connect adjacent photocells to threshold devices with a delay, we have a motion detector specific for a given direction and velocity.

We can also use lateral inhibition to enhance visual detection.

All of this requires some orderly description of sensory space. This also enables us to do spatial computations. If we have 2 points in space that are represented, we can calculate a direct path between them. This is done by imagining a sheet that has a uniform conductance, and applying a voltage between two points on the sheet. The current flow will be along the straight line between the points.

A continuous representation of space allows for continuity of movement to be detected, and this is a primary criterion for the physical reality of an object.

Also, the representation of space can be of many kinds (e.g., audio, theoretical, etc.) and be in many dimensions above 2.

Vehicle 9 – Shapes

A detector for bilateral symmetry. The vehicles can now recognize vehicles, but must also take into account their behavior – i.e., that they are facing forward. This provides information about “being in someone’s focus of attention.”

There could also be detectors for radial symmetry, which indicate singularities in the world. A fundamental characteristic of things in the world is periodicity, so we should equip our vehicles with detectors for periodicity. This can be done with periodic templates of different spacing and the mathematical technique of cross-correlation and Fourier analysis. Note that something like this seems to be how the hippocampal formation represents space.

Vehicle 10 – Getting Ideas

The claim here is basically that a vehicle of type 7 can get “ideas” – this is nothing more than more complex concepts, based on statistical relationships. For instance, if there are rows of flowers, and edible flowers exist only in rows 1, 8, 15, 22, etc., then the concept of the number 7 may emerge.

Vehicle 11 – Rules and Regularities

We introduce a new way of learning – not just simultaneous associations but causal effects. Introduce a new type of wire – Ergotrix. This wire acts like Mnemotrix, but it is directional. It operates only when event A precedes B, and not vice versa. We will start by connecting ergotrix in both directions between every threshold device. We will also make the condition that whenever the Ergotrix wires get strengthened, the Mnemotrix wire between each group become strengthened too.

These vehicles would eventually display particularly well defined reactions to events that are known to have consequences.

Vehicle 12 – Trains of Thought

Let all the thresholds on the threshold devices be controlled by a universal controller (could be more localized). When the overall activity in the brain is high, the thresholds are set higher, and vice versa. This allows the vehicle the FOCUS ATTENTION on the thing with the highest activation. But since we also have Ergotrix and Mnemotrix, related concepts would then get activated, and a continuous train of thought could occur.

The number of active elements at time t+1 as a function of t might look something like a logistic map. In this case, the transition sequences would be highly unpredictable, which imbues our vehicles with the appearance of free will.

Vehicle 13 – Foresight

What we want is some semblance of purpose guiding the behavior, so we can have goal-directed behavior. Two important aspects of goal-directed behavior. 1) The goal lies in the future, and 2) the goal is desirable.

The action is a consequence of something we expect to happen in the future. What we need is a mechanism to predict future events fast enough so that they will be known before they actually happen. We then need stored sequences of events, plus a mechanism forcing them to speed up in the reproduction when necessary, e.g., dangerous situations. We also want the ability to consider several likely predictions in parallel.

This is a highly complex chapter, with serious considerations on how to develop a brain that not only can use predictions, but knows when to rely on past predictions and when to rely on sensory input. This mechanism is driven in no small part by the connections forged by evolution. We also give weight to rare but important events.

Vehicle 14 – Egotism and Optimism

New rule: When choosing among several equally likely next brain states, choose the most pleasing one. The Darwinian evaluator cycles through the possible predictions and evaluates them for favorable or unfavorable aspects. This generates certain values for each prediction, and we choose the one with the highest value.