Active Semantic Networks In Robots With Independant Control

Introduction

Considerable specifics is inherent in problems of constructing the control sys­ tems of robots designed for functioning in the natural environment. Here elaboration of such decision procedures proves to be necessary which can operate not only under conditions of unsteadiness and variability of initial data of the problem being solv­ ed but also under conditions of simultane­ ous existence of many problems whose ur­ gency is in constant change. Hence the stringent requirements that the robot con­ trol systems be versatile and autonomous. Moreover, they must function under conditi­ ons of real time and considerable deficit of information. Approaches based on the traditional methods of decision-making, . theorem proving, etc., seem to be not very effective. To the best of our belief the promising direction is associated with the development of such decision-making systems which are based on the employment of mecha­ nisms of "inner activity" in the sense of the term meant when speaking about the motivation of human behaviour. It implies, however, the development of a method fit for the new tasks.

This paper offers the description of a possible method of the kind. A certain elaboration of the idea of semantic net­ works forms its basis. Ways of the method application in constructing control sys­ tems for robots are discussed and examples

of specific developments are given.

Sign Systems Ind Information Va ue

The Chosen Approach

It is characterized by two main fea­ tures. 'l'he first: we construe t the robot control system on the basis of knowledge and hypotheses on processes of information handling by the human brain which affects the behaviour we would like to observe on

25262? Kiev USSR

the part of a robot. Initial hypotheses are given in [11· It does not at all mean that we shall be engaged in the modeling of the thought or other psychic functions. It is only required that the constructed system control the robot efficiently. We use the study of psychics of an individu­ al affecting the control only as the heu­ ristic basis for incorporation of one or other functional subsystems or structures into the system.

The second feature consists in our employment of the semantic network, trea­ ted as a sign system, as a device for the internal representation of information in the system. ·

In works of different authors the semantic network appears in quite differ­ ent modifications, i.e., from facilities for representing meanings of lexical units in [21 by R. Quillian to "a state-of-the­ world graph" in [ 31 by G. Hendrix. The semantic network we deal with is an orien­ ted graph whose nodes and arcs are marked by names of some notions. The network no­ des are considered as the signs whose de­ notata (i.e., objects corresponding to them, actually existing in the real world) are both objects and phenomena of the en­ vironment as well as images, notions, feelings, emotions, etc., existing in the inner, psychic world of an individual be­ ing a prototype of the constructed control system. 'rhe network connections are treat­ ed as the signs whose denotata are rela­ tions actually existing among the denotata of nodes or among their concepts . Thus, the representation of interactions among essences quite different and 1'ar 1'rom one another becomes feasible in a single sys­ tem and in a sing~~ language. Each sign is put in correspondence with a designatum, i.e., a notion or concept of an actual ob­ gect. Nodes and connections of the seman­ tic network are marked just by the names of these notions. What may be said about syntactics and semantics of our system is more or less obvious. Let us dwell on its pragmatics.

Pragmatic Closure

In the general sense pragmatics in­ cludes the set of relations between the sign system and its user or of sign-inter ­ preter of the sign relations. Our sign system is, so to say, "a brain" of the de­ vice designed for autonomous functioning . Therefore, it is also an independent con­

trol system. This creates a particular situation when the sign system is, strictly speaking, its own user, its own interpre­ ter, since it or~anizes its behaviour em­ ploying the sign forms of representation that exist in itself, ~his circumstance lends distinguishing features to the prag­ matic aspect of the study of such systems. To emphasize them we define systems of that kind as pragmatically closed in con­ trast to the traditional systems - prag­ matically open - where the sign system is something different from its user. It is, apparently, reasonable to consider the sy­ stem of human thought as the pragmatically closed one. We believe that independent systems of artificial intelligence must be inevitably the pragmatically closed sys­ tems.

One of the most interesting problems treated within limits of pragmatics is that of value.

Information Value

The notion of value is complex and multicomponent. However, if the study is carried out in comparatively narrow bounds of the field of pragmatically closed arti­ ficial systems, then it turns out to be possible to limit (to begin with) a volume of the information value notion including in it, at the first place, such components as significance of atoms of information, a degree of their involvement into the de­ cision-making process, and a degree of their interconnection. Here it is of im­ portance that information value characte­ ristics should not be specified to the sy­ stem beforehand but be determined by the system itself in the process of function­ ing with due account of its own state, ex­ ternal situation, and "motives", i.e., "subjectively". To solve this problem we introduce the notion of semantic network activity.

Representation Method

Activity in Semantic Network

Let us assume that in an environment, for instance, a stone lies on the way of a robot, that the latter percieves and re­ cognizes it. If the robot "knows" what a stone is, then a robot marked by the cor­ responding word should exist in its seman­ tic network. Let such a node exist. Thus, the stone is on the way, it can hamper the movement, it is the reality. Accordingly, i nformation about it, hitherto stored in the memory passively, becomes significant. In the semantic network the node STONE is a "key" to this information. 'l'his is the node that must be actively used by the control system now: the reality of the stone must be transformed into a rise of a certain activity, into a growth of the value of notion of a stone. A means of this process representation used by us consists in assigning some value of a specific ana­ log quantity - an excitation - which is a figure of significance or, in the general case, of value of the corresponding atom of information, to t !te necessary noue.

We shall assume that the assignement is performed by the specific perception procedure whose way of realization will not be considered here. Thus, asswne that the perception procedure has ensured some excitation of the node STONE, has escited it. One node is connected with many others in the semantic network. Let there be a connection marked by the name BELONG TO CLASS relation that goes from the STONE to the node OBS~ACLE. The robot actually perceives only the stone. However, it is quite natural to require that the seman­ tic memory give it the possibility to "understand" correctly the meaning of what is being perceived and enable it to ope­ rate with this meaning. In other words, if the atom of information STONE is sig­ nificant the atom of information OBSTACLE must also become significant, as though the V~L~e of the first induces the value of the second through the connection exi­ sting between them.

It is possible to reason similarly about other connections of the node STONE in the semantic network. Thus, the connec­ tion IS A PART may connect this node with the nodes MOUNTAIN or, say, SCllliE. And though the mountain or scree themselves might not yet be perceived by the robot, the value of corresponding nodes should be induced by the STONE and that will ensure the activity of the relevant information and its participation in the decision-ma­ king. 'l'he same is with connections of the type ENTAILS (for instance, HAIN-ENTAILS­ BAD PASSABILITY OF SOIL), HAS, etc.

A constructive representation of the value inducing process discussed is the process of excitation transfer along a connection .in such a way approximately as it occurs in electric circuits or neuron structures. Here the connection has a cer­ tain wei~t characterized by some number. The w~ig is an analog quantity playing the part of the coefficient of a decrease or increase of excitation transfered along the connection or, in other words, the part of the index of correlation between value characteristics of connected nodes . Such indeces can be not only positive but also negative (as, for instance, in the string BAD STATE OF SOIL - HAMPERS - RAPID MOVEMENT).Connections may be reinforcing or inhibiting, respectively.

Therefore, excitation of even one node can result in excitation of many others connected with it directly. From these others the excitation can propagate still further and, as a result, the value characteristic can be obtained by very ma­ ny, in the general case by all, atoms of information stored in the network.

Activity Sources

Let us recall that we construct our semantic network not only as a system for storing the robot's knowledge on the world but also as "a model" of some indi­ vidual being a prototype of the control system constructed. It enables us to in­ clude into the network fairly specific notions, such as notions of feelings, de­ sires, aspirations, etc. The general scheme of reasoning in 1..heir derivation looks approximately as follows.

If problems that a robot faces were tackled by a ~an then in the given situ­ ation such and such emotional state (or an aspiration, an image, etc.) would oc­ cur in him. It follows from such and such data of the psychology (or obtained by us during a specific experiment, or, finally, it is obvious intuitively). Very good. Let us introduce a sign of this state in­ to our network, mark it by a name of the notion of this state, and connect this sign (i.e. the node) with other node.s of the network in such a way that is based on psychology ( an experiment, intuition) data.

Such a scheme permits to assimilate into a single system a set of knowledge of different kinds on psychics - the most efficient of control systems existing in Nature. This scheme does not, however, compel us to imploy in succession all such knowledge and nothing but it.-wi need only that is actually useful in con­ struction of an efficient system.

The described method application allows in an active semantic network to represent structures corresponding func­ tionally to the structure of human moti­ vation mechanisms and including notions that describe needs, aspirations, sets towards smth., etc.

The nodes corresponding to such no­ tions can be excited in the network pri­ mordially and their excitation value can change compa:ratively slow. It is, so to say, the absolutely valuable information determining the global set of the systems. From them excitations propagate on all other nodes of the network giving them some initial "set" value. 'l'his initial distribution is superposed by another constantly changing distribution generated by work of the perception procedure and reflecting situations in the environment. Both these processes unuergo a constant mutual influence. Thus, a sequence of some specific perceptions can chanBe the exci­ tation of nodes corTesponding, for insta­ nce, to feelings and influence the activi­ ty of one or other motives through them. Such a process of mutual influence deter­ mines "the subgectiveness" of information value characteristics.

One of the activity regulators are the network inhibitory connections whose influence on each of the nodes opposes constantly to the exiting influence of the reinforcing ones. A trend, realized by functional characteristics of the node, to a spontaneous decrease of its own excitation in absence of input effects ("the memory") is another regulator. How­ ever, the special System for Reinforce­ ment and Inhibition (SRI) is the basic mechanism of network activity regulation.

The SRI principal purpose is to pro­ vide the domination of one atom of infor­ mation over all others at each time in­ stant. SRI affects the positive excitation feedback in the network giving an additi­ onal excitation to the node most active at this moment and inhibiting all the others proportionally. SRI - is a dynamic system constantly switching its extra ac­ tivity from one node to another changing here the level itself of an additional excitation and inhibition.

The SRI construction reflects the adopted by us ideas about the attention mechanisms structure whereas the SRI role in network operation is compared with that of attention in the thought process. This comparison serves as one of the instru­ ments of both the analysis of processes taking place in the network and the syn­ thesis of the network itself in that part of it which is based on psychology data.

An important role in the network ac­ tivity regulation belongs also to the learnine and self-organization processes securing a constant Change of weights Of the connections available, occurrence of new connections, change of forms of func­ tional characteristics of nodes (their "memory", in particular), as well as "a growth" of the network on account of occ­ urrence of the new nodes corresponding to new notions that are formed in the network. Development of all these proces­ ses is determined by a character of final results of behaviour generated by the net­ work and is directed to keeping an opti­ mal relation of excitations of the spe­ cial nodes PLEASANT and UNPLEASANT reflec­ ting the general estimate of state of the motivation sphere.

Decision-Making

The method of building, employing and the composition of the suggested net­ work are sufficiently specific to justify introduction of a special term for the de­ signation of semantic networks with an activity of the decsribed type. Let us call them M-networks. The M-network is only a device for solving particular prob­ lems. Specific control systems built as M-networks and possessing SRI are defined as M-automata.

Let a robot control system be con­ structed as the M-automaton. Consider basic processes taking place in it during its functioning.

The perception procedure transfers into the network information about a current situation of an environment by the exciting of co~responding nodes. It is obvious that not all M-network nodes may be excited by this procedure. Those of them whose excitation ls controlled by the perception procedure are called the receptor ones. Thus, the node STONE was a receptor one in the above example.

From the receptor nodes the excita­ tion transfers into the network where mixes with excitation going from perma­ nently active nodes representing inner motives of the system. Value characte­ ristics of all atoms of information stored in the M-network are formed as a result. The process of excitation propa­ gation is controlled by SRI. A certain imo~e, a notion, a feeling, etc., is "singled out by the attention" and domi­ nates at each time moment. It is, so to say, "a thought" of the system. Continu­ ous switchings of SRI form a flow of such "thoughts", while the switchine;s them­ selves appear as associative transfers. Nodes that are not singled out and are, accordingly, inhibited by SRI compose "a subconsciousness" of the system to which additional excitations inflow con­ stantly along connections from the nodes singled out by SHI and where a continuous "competition" takes place 1'or the "atten­ tion" capture. In the network during the­ se processes stable excitation centers come into being, exist a certain period, and disintegrate, which can be also slu­ ggish, and latent, and "realizable". From time to time the flow of these processes is perturbed by a new portion of excita­ tions brought into the network by the perception procedure when external situ­ ations change. Therefore, the distribu­ tion of excitations in the network is dy­ namic, i.e., continuous correction and reestimation of the value of information available is performed in the "thought" process.

There are nodes in the M-network which cor-respond to notions of those real actions that a robot is able to car·ry out in an environment: EXECUTE 'rURN TO THE LEFT,TAKE OBJECT X, etc. Let us define such nodes as effectory. There are also nodes correspond~ng, so to say, to gene­ ralized actions, e.g., AVOID ENCOUNTER WITH OBJECT Y (no instructions here how to avoid). As a rule, connections with greater weight are established from nodes of that type to the effectory ones. If SRI single out an effectory node or node of generalized action in the process of M-automaton functioning and if such node begins to dominate over the others we say that the system has made an appropriate solution. If SRI singles out an effectory node it is interpreted by the system as a command that the corresponding real ac­ tion be executed by the robot. We shall assume that the corresponding action is executed by some execution procedure. Depending on the specific robot construc­ tion and specialization this procedure may have an access to subsystem of direct control of its actuators, manipulators, etc.

Thus, the general structure of pro­ cesses realized by the M-automaton is as follows. The perception procedure ref­ lects information about an external si­ tuation into the M-network. The process goes of reestimation available in the network. Checking of the environment state is performed during this process. A decision is made at some instant and the execution procedure realizes an appropriate action. It results in a change of environment, the new situation is reflected into the M-network, and the whole cycle is repeated. During multiple repetition of such cycles the robot's behaviour is being formed.

Specific Developments

Hefinement and investigation of the described approach was carried out by creation and study of possibilities of a number of functioning M-automata realized as digital computer programs. As an object of modeling the information handling pro­ cesses performed by a man when organizing a purposeful motor behaviour in static natural environments were selected. On the one hand, these processes are sufficient­ ly complex to ensure the check of "a strength" of the applied methods, on the other hand, a number of interesting prac­ tical problems - in particular the cont­ rol of robot motion - are reduced to the problem of purposeful motion organization. The developed M-automata imitate actions of a robot in an environment with objects that are to a various extent useful or dangerous to it. Motives of the robot behaviour depend on its striving for self­ preservation and achievement of a global goal. The robot studies the environment, chooses a motion objective, constructs a plan for its achievement, and then ful­ fils it, performing actions-steps and actions toward transformation of the en­ vironment, comparing obtained results with the planned ones, amplifying and correct­ ing the plun in accordance with existing situations.

The detailed description of all de­ veloped M-automata is eiven in the mono­ graph [43 and works (5,6]. There are deve­ lopments of technical realization of the robot control system [ 5, 7 J.

M-Networks and Neuron Ensembles

There exists one more independent way of the suggested method argumentation.

It is associated with treatement of M-net­ work nodes as elements corresponding to ensembles of neurons in cerebral cortex. Such ensembles, the idea of which ascends to Hebb ra), are treated here as cortical information models of elements of an In­ dividual's inner and external world. In this case the M-network appears as an apparatus for the description of an in­ tegral inner (cortical) model of a world of a man and the M-automata as psycho­ physiological and psychological cognition models. A number of our investigations is l ompleted within a such interpretation 1,9). Simultaneous existence of two in­ ependent to such an extent processes is, in our opinion, quite an interesting fact. It allows , in particular, to suggest (not but speculatively for the time being) that processes of activity of cells and greater brain formations studied by neurophysiolo­ gists may be described adequately as pro­ cesses of execution of procedures of sub­ jective estimation and reestimation of elements of the inner world, i.e., pro­ cedures which underlie the behavioural reactions formation. When developed a sudl approach would serve the cause of narrow­ ing the gap between sciences studying the brain and behaviour.

Conclusion

Semantic networks are traditionally employed to represent a system knowledge on the world in artificial intelligence systems. We endeavoured to use them in representing our own 'knowledge not only on the world but on mechanisms of beha­ viour formation, too. Having preserved the constructivity of information presentation we gained an opportunity of using the de­ veloped apparatus in artificial systems, in robots in particular. Having introduced the notion of semantic network activity we obtained the possibility to imitate the inner activity or motivation of a robot. Processes of motives-knowledge interac­ tion were represented in the dynamics of information value characteristics. Comp­ lexity and diversity of behaviour forms which were observed during the experimen­ tal study of the developed M-automata con­ vince us in the efficiency of the used approach and in expedience of its practi­ cal empl oyment in efforts to develop in­ tegral r obots. This may offer great possi­ bilities of representation of complex and intimate behaviour mechanisms knowledge on which may be drawn from the psychology.