Synthetic psychology. Recognition and classification. Similarity measures.

Synthetic psychology

In the photo is Gray Walter's bug


"The book of Valentino Breitenberg, inspired an entire generation
robotics and cognitive scientists, has become just a classic.
She will change your understanding of the psyche "
D. Dennett. ("Types of psyche")



Synthetic psychology is a new trend in psychology and cybernetics, which, through the use of elementary means, makes it possible to imitate complex biological behavior.
In 1984, Valentino Braitenberg (Valentino Braitenberg), a famous Italian neurologist and cybernetics, professor of the Institute of Biological Cybernetics. Max Planck in Tübingen (Max Planck Institute for Biological Cybernetics), published a small book: "Cars: Experiments in Synthetic Psychology" (V. Braitenberg "Vehicles: Experiments in Synthetic Psychology", A Bradford Book / The MIT Press, 1984), which inspired a whole generation of robotics and cognitive scientists. Breitenberg, using his principle that "biological behavior is easier to synthesize than analyze", made from elementary mechanisms and electrical devices 14 variants of synthetic models (machines), controlled by simplest schemes that imitated complex behavior (love, aggression, fear and foresight). In this book, the author provides data surprising behavior of his machines. The book has become a bestseller in the field of SYNTHETIC PSYCHOLOGY.
In 1962, on the basis of a self-propelled electric caterpillar toy, he assembled a cyber: a machine that possessed the ability of phototropism (striving for light). To do this, in the front of the installed two sensor-photoresistor, connected through the pavement electrical circuit with opposite electric tracks. In a dark room, the self-propelled gun itself found a source of light and rushed towards it. I would not have remembered this long-standing passion if it were not for one surprising coincidence - my cyber was the simplest version of the phototraction of Breitenberg machines!

It should be noted that much earlier Valentina Breitenberg, English neurophysiologist Gray Walter (William Gray Walter) from 1948 to 1951 created biorobots ("turtles"). Gray Walter called them machina speculatrix. The "turtles" were self-propelled electromechanical carriages capable of crawling into or out of the light, bypassing the obstacles, entering the "feeder" for recharging discharged batteries. On the steering column was installed photocell and mechanical contact, closed when hitting an obstacle. The behavior was controlled by two simple artificial neurons. In total, Gray Walter built eight turtles. Thus, he has already shown that simple turtles had a complex behavior of biological systems (search for light ("hunger"), turn to the light, avoid bright lights and obstacles ("pain")). Therefore, Gray Walter and Valentino Breitenberg, with good reason, can be considered the founders of SYNTHETIC PSYCHOLOGY.
The very idea of ​​synthetic psychology was first put forward by the French psychologist Alfred Binet (1857 - 1911). He is also widely known for creating the first psychometric test similar to IQ. The famous Swiss psychologist Jean Piaget (1896-1980) and the American psychologist Mark Baldwin (1861-1994) were involved in the ideas of synthetic psychology.

Sometimes synthesis is easier than analysis. It is difficult to understand how the brain works at a low level, but you can try to imitate behavior as <input-output>. The purpose of this direction is the creation of behavior simulators. Such imitators were created by various groups of researchers - Amosov, Breitenberg, Walter and others. At the lower levels, machines imitated the simplest emotions and behavior (fear), at the upper levels, more abstract and complex (premonition, love). On the simplest model, you can see interesting psychological reactions.

Consider the “Turtle” by G. Walter, which he called “The Speculator Machine”, which means the “thinking machine”. "Turtle" is a three-wheeled cart, which houses the batteries, relay system and electronic (lamp) circuit. The front wheel is located on the steering column and is driven by two engines - steering D ₁ and driving D ₂ . On the steering column is also fixed photocell and a light bulb, the light from which does not fall on the photocell. The cart is covered with a casing resembling a turtle shell in shape and leaving only a photocell and a light bulb open. Around the housing casing is an elastic bumper, the touch of which closes the mechanical contact K ₁ .

The schematic diagram of the "Turtle" is shown in Figure 12.1.

Fig. 8.16. Electrical circuit diagram of a robot turtle

A “turtle” with such a scheme is capable of reproducing the following types of behavior: movement into or out of the light, search movements, bypassing an obstacle, movement towards the “feeding trough” and entering into it to recharge batteries. The circuit is adjusted so that when the anode potential of the lamp L _{1 is} low, the lamp L _{2 is} locked and the relay P _{2 is} set so that the simultaneous presence of the relay P ₁ and P ₂ under current is excluded. If the PV cell of the PV is not illuminated, then the lamp L _{2 is} open, and L _{1 is} locked. Under moderate photocell illumination, the L ₂ “opens up” lamp, however, the current it conducts is insufficient to trigger relay P ₁ , although a decrease in the voltage at the anode of the lamp causes the release of relay P ₂ . A further increase in the illumination of the photocell (“blinding”) leads to the triggering of P ₁ when P _{2 is} released. The closure of the mechanical contact K ₁ turns the circuit into a multivibrator, alternately switching on and off the relay P ₁ and P ₂ .

The behavior of the "Turtle" depends on external influences, i.e. from the relay states and is characterized by the following table.

Despite the relative simplicity of the device, "Turtle" demonstrates quite diverse types of behavior. Under normal conditions, it carries out a “search”, moving along the trajectories it forms, and for an hour is able to explore a rather large room. When light hits the photocell, it either approaches the light source (in low light or dead batteries), or moves away from it (in strong “blinding” light). When hitting an obstacle - no matter what other stimulus acted - the machine only responds to touch, which is a stronger stimulus, and moves away from the obstacle. After the waste maneuver, he continues to move under the influence of the remaining stimulus. "Turtle" is able to find the optimal condition for itself, which is the search in dimly lit places. The lamp L ₃ , located in front of the device, allows it to “recognize” itself in the mirror in the absence of other light sources, approaching its image. In the same way, two machines can "see" each other and move towards. When batteries are partially discharged, the ability to avoid “blinding” weakens and the Turtle approaches the “feeder” in which the light source is installed, and when it comes into contact with the contacts of the charger in it, “saturates” and then (“blinding” occurs) moves away from her and continues to normal activity. Below are examples of different behavior of the "Turtle": search in the absence of bright light, the desire for a not very strong light source, the behavior in the presence of two strong sources of light, the avoidance of an obstacle with the desire for light, the desire for one light source of two identical automata, a visit to the "feeder ", the behavior in front of the mirror and the" acquaintance "of the two" Turtles ". It is easy to see that the behavior of the automaton is in many respects analogous to the complex of fixed actions observed in the behavior of living organisms. Note that it is impossible to establish the exact trajectories of movements in advance. In each case, they are determined by a number of random factors, such as the roughness and inclinations of the surface along which the machine moves, the friction in its moving parts, the scatter of the relay response times, etc.

About machines Breytenberg - machines with synthetic psychology, you can read more details ...

A total of 14 levels of such animals were built. At the third level, we managed to implement the predator-prey scheme. On the fourth - the machines distinguished living from nonliving, skirting obstacles. On the fifth implemented the simplest training. At 6-8 levels learned to generalize. At 14, they developed a scale of values, found a line to achieve the goal as quickly as possible, and created an internal model of the world.

A significant drawback - there is no self-production and evolution.

Recognition and classification

The task is to separate a sample of type A from a sample of type B, for this it is required to enter a physical quantity that would distinguish samples of different types. Obviously, this value will have a whole range of values ​​and choosing a certain identifying attribute we can face the overlap problem (upper graph), that is, the situation of the presence of this feature in both samples. Therefore, it is required to unfold the base axes so that the projections of the various classes intersect minimally (lower graph).

The task of intelligence is to find the lines of class division. For better results, additional features are required. Each is useful for at least one division and affects the reduction of the likelihood of error. The best signs are those in which the difference in averages is higher than standard deviations. A sign will be useless if the averages coincide with the averages of another sign. A lot of signs increase the likelihood of error.

Similarity measures