WO2014195633A1

WO2014195633A1 - Analysis of complex systems

Info

Publication number: WO2014195633A1
Application number: PCT/FR2014/051322
Authority: WO
Inventors: Norbert Rimoux
Original assignee: Norbert Rimoux
Priority date: 2013-06-04
Filing date: 2014-06-04
Publication date: 2014-12-11

Abstract

This method of analysis of a complex system, implemented by digital processing means, comprises: • - receiving a digital data stream; • - subsequent to the receipt of a particle consisting of a fragment of said stream, creating and storing a computerized object associated with this particle (22), said object comprising an identifier (pID); • - detecting strong interactions between said particles, • - determining for each particle, a segmentation index value (IS) representative of the number of strong interactions detected for said particle • - selecting at least one particle as segmenter of the stream as a function of its segmentation index value, • - creating composite particles consisting of aggregates of particles (26), on the basis of the particles received in the stream between two segmenters; • - using these aggregates to hierarchize and categorize the data of said stream.

Description

ANALYSIS OF COMPLEX SYSTEMS

The invention relates to the analysis of complex systems, that is to say the structure of which is unknown or unstructured. It can find an application in any delicate or impossible to model domain, for example human language, complex physical systems (meteorology, three-body systems, molecular interactions ...), DNA chains, etc.

The invention aims to determine a structure of this complex system. Pre-structured or mathematical modeling data can be relatively easy to organize and use. For example, data can be stored in database table fields, or computer modeling software can be used. On the other hand, the processing of unstructured data is more difficult. Various methods can be mentioned, including statistical methods, for example Markovian or Bayesian processes, methods based on neural networks, or other methods. These methods generally involve human supervision. EP 1 306 768 discloses a pattern recognition method, for example words, from a received sequence of events, for example letters in an ASCII format. When receiving an event, a signal element, or cell, is activated. Some cells, preceding a predefined event, such as space, comma, or period, are identified as ending a word.

There is a need for more relevant segmentation.

There is provided a method for analyzing a complex system, implemented by technical means, for example by digital processing means, comprising: providing an initial set of particle elements, receive data forming a sequence of events produced by the complex system,

following reception of an event of said sequence, activate / identify a particle element then associated with said event,

detecting strong interactions between the activated / identified particle elements, each strong interaction corresponding to two particle elements for which the event corresponding to one follows the event corresponding to the other in the received sequence of events,

determining for at least one, and preferably each, identified / activated particle element an index value from the possible strong interactions detected for this particle element,

selecting at least one particle element as a segmenter according to said index value,

detecting events corresponding to segmenters in the received sequence and creating aggregates of particles, each aggregate corresponding to a sequence of events received in the sequence between two events corresponding to segmenters. The invention thus provides a method for analyzing a complex system, implemented by digital processing means, comprising:

receiving a stream of digital data produced by the complex system, following the reception of a particle constituted by a fragment of said stream, creating and storing a computer object associated with this particle, said object comprising an identifier;

detect strong interactions between these particles, each strong interaction corresponding to two consecutive particles in the stream, storing in a so-called "strong causal coupler" register of the computer object associated with each particle, the identifiers of the computer objects of the other particles having at least one strong interaction with this particle; determining for each particle a segmentation index value representative of the number of strong interactions detected for said particle and storing said index value in the computer object associated with said particle ",

selecting at least one particle as a segmenter of said stream when the value of the segmentation index associated with this particle exceeds a threshold value,

- Create composite particles consisting of aggregates of particles, each aggregate corresponding to a sequence of particles received in said stream between two particles corresponding to segmenters;

use said aggregates to prioritize and categorize the data of the stream.

Thus, and in a general manner, the invention makes it possible to prioritize and categorize the data of a digital flow without any prior knowledge on this flow.

In the case of written language, for example, the method described above can make it possible to identify the main segmenters of the language, then the possible structures constituting a signal of syntactic type, then the words, the phrases, and their contexts. membership. For example the word "black" can belong to:

- a common context with "neutron star", "astronomy", "black hole", etc.

- another context with "color", "red", etc. , and

- another context with "skin", "civic rights", "Martin Luther King", etc. In a gene detection application, the knowledge of the segmentation information can make it possible to discriminate the position of the gene to be searched more correctly than in the prior art.

In a particular embodiment, the method according to the invention comprises a step of using the segmenters to minimize the memory required for encoding the data of said stream.

For example, still in the domain of language, the expressions:

- "chairman of the board" comprises 3 particles; - "Chairman of the Executive Board" adds 1 particle;

Therefore, and following the same logic, the digital flow of "chairman of board and directory" comprising 7 particles can be encoded on 5 particles.

The invention thus very advantageously makes it possible to compress a stream of digital data over which there is no prior knowledge. It obviously finds a preferred but non-limiting application in search engine applications and more generally indexing language or any other type of information. In a particular embodiment, the method according to the invention further comprises: re-ordering said identified particles in an ordered sequence, according to the segmentation indices (IS) of said particles; detecting weak interactions between said particles, each weak interaction corresponding to two consecutive particles in said ordered sequence, the segmentation index value stored in the computer object associated with a particle being representative of the number of strong and weak interactions detected for said particle, storing in a so-called "contextual coupler" register (12) of the associated computer object to each particle, the identifiers of the computer objects of the other particles having at least one weak interaction with said particle creating a context with at least two particles the infomatic object associated with each of said particles comprises in its contextual coupler register the identifiers of the computing objects of each of the other particles; using said contexts to generate and store a compressed version of said data stream.

Thus, it will be understood, also by way of nonlimiting example, that the invention may be used in applications aimed at creating text summaries dealing with topics that are a priori unknown, or even in a language that is a priori unknown.

Thus, the particle elements, or particles, are detected preceded or followed by a large variety of particles. The segmenters thus obtained are more relevant than if they were predefined, as in the prior art.

This method can moreover be automated relatively easily.

In addition, this method can make it possible to identify segmenters whatever the nature of the data received. In the case of events received in the form of an ASCII signal, it can indeed be assumed that the character ESPACE plays a role of segmenteur between the words, but to define manually and in advance the segmenters can be delicate when the data received are data of a different nature, eg audio data, position values, etc. The invention can thus make it possible to detect the most relevant segmenters, while avoiding the nature of the data received.

The particles of the initial set can advantageously be undifferentiated.

The initial set of particles may advantageously be finished. Thus, the number of identified particles is limited, which can make it possible to carry out the method with reasonable memory and / or calculation resources. Advantageously, provision may be made for a possibility of particle identification, corresponding for example to a return of a previously identified particle to an undifferentiated state.

Thus, initially, the particle elements can be in some ways compared to physical particles in a free state, not discernible, non-materialized, the number of particles being finite but possibly relatively high. For example, this initial set of contextual voids can be qualified by analogy with the quantum vacuum.

The invention can be implemented by digital processor processing means.

The digital processing means may for example include or be integrated in one or more processor (s), for example a microprocessor, a DSP (the "Digital Signal Processor"), or other. The particle elements may be computer objects, that is, memory is allocated to store data, the allocated memory being structured to associate data relating to the same particle element with each other. . The memory allocation can be dynamic or static: for example, N memory spaces are provided, each memory space corresponding to a particle element, or even a number of particles is stored. undifferentiated, and during the materialization of a particle element dynamically allocates a memory space for storing data relating to this particle element.

The invention is however not limited by an implementation of the invention by digital processing means. It is not excluded that the invention can be implemented by analog means, such as the system of real particles in interaction, photon traps, or other.

Nor is the invention limited by the nature of the complex system, nor by the nature of the sequence of events produced by this complex system. For example, the complex system may be a three-body system, and the event sequence a series of spatial coordinate values for each of these bodies. In another example, the complex system may include a language, and the sequence of events may include a spoken audio signal, ASCII text, or the like.

When an event is received and there are no previously identified particles corresponding to this event, it may be possible to identify a new particle for this event. If there is already an identified particle for this event, that particle is activated. In both cases, one can then examine the particles preceding / following in order of the sequence that received particle to better estimate the level of meaning of this event, a relatively meaningless event being relatively likely to be selected as segmenter.

Advantageously and non-limitatively, during the selection step, the index value of a particle element is compared with a threshold, and said particle element is selected if said index value is greater than said threshold. In other words, particles that can be preceded and / or followed by a large variety of particles are chosen as segmenters. For example, particles corresponding to punctuation marks of the type ",", ". Have a relatively high chance to be selected as a segmenter because these particles are likely to be preceded / followed by most other particles.

On the contrary, particles which are found to be preceded and / or followed by a small number of particles may be more meaningful. For example, in the case of an event sequence comprising an ASCII character sequence, the particle associated with the event "B" may often be followed by the particle corresponding to a vowel or "R" events, "L" for a sequence in French. The particle associated with this "B" event is likely to have a relatively low chance of being selected as a segmenter.

Advantageously and in a nonlimiting manner, it is possible to re-order the events of the received data which are associated with identified particle elements in an ordered sequence, as a function of the indices corresponding to said particle elements, for example in ascending or descending order. of these indices.

Advantageously and in a nonlimiting manner, it is possible to detect weak interactions between the activated / identified particle elements s, each weak interaction corresponding to two particle elements for which the event corresponding to the one follows the corresponding event. to each other in the ordered sequence.

Advantageously and in a nonlimiting manner, the threshold may be chosen equal to Pe ² / Cx, Pe being the number of particle elements identified among the set of particle elements identified in the stream, and Cx being the number of strong and / or weak interactions detected between said identified particle elements and associated with said events.

Surprisingly, this threshold value can make it possible to obtain a relatively stable process with relatively relevant segmentors. Alternatively, one could of course choose another way of calculating the threshold.

Advantageously and in a nonlimiting manner, the index value of a particle element can be determined according to any weak interactions detected for this particle element, for example as a function of the interactions for which the event corresponding to this particle element. particle element follows and / or precedes another event in the ordered sequence.

For example, the index of a particle element may be a function of the number of interactions of which that element of particle is the cause. A particle corresponding to an event that has been observed to be preceded / followed by 20 different events may thus have a higher index than an event that has been observed to be preceded / followed by only 3 events. We can plan to take into account the number of activations.

For example, an event received 10 times and for which 10 different interactions have been identified may have a higher index than an event received 100 times and for which 10 different interactions have been identified. The invention is in no way limited to such a consideration of the occurrence frequency of the event. For example, it can be expected to keep only the events received relatively often, so that relatively rare events can see their associated particles disappear and eventually reappear, the corresponding index thus remaining relatively low compared to indices corresponding to more events. for which the particles remain more stable.

Advantageously and in a nonlimiting manner, the index value of each particle element is compared with a low threshold, and the particle element is kept among the identified elements only if said index value is greater than this low threshold. . Thus, infrequent and / or low interaction particles are removed from the identified group of particles. An eliminated particle returns to the free, undetectable state of the initial set of particles. In one embodiment, this threshold may change with the number of advent events received and / or the number of interactions detected.

Advantageously, a method based on buffer-type storage will be chosen, in which, if, during the last N identifications / activations, a particle previously identified has not been activated again, its index value decreases. In which case, the threshold may be fixed, for example zero.

Advantageously and in a nonlimiting manner, for each aggregate of particles created, a new particle element is associated with said aggregate and an index value is determined for said particle element associated with said aggregate.

In other words, aggregates can be treated as particular particles. The method can thus make it possible to manipulate groups of consecutive events rather than isolated events.

Advantageously and in a nonlimiting manner, when a set of at least two, and preferably at least three, particles of particles are in weak interactions with each other, for example if this set of particle elements forms a loop causal, then one can transmit a context detection signal to activate / identify a particle element associated with this context. Particle elements identified and associated with events of the received sequence can thus form a first stratum. If it turns out that the particle elements of a set of particle elements of this first layer form a causal loop, ie if a context has been detected, then a particle of another stratum, called higher level, is identified / activated. In other words, Γ identification / activation of a context within a stratum leads to the identification / activation of a particle within the upper stratum. The different identifications / activation of particles in this higher-level stratum form somehow an incoming particle sequence in this higher-level stratum, within which strong and weak interactions can be detected, in the same way as for the first one. stratum except that the data received is no longer a sequence of events from the complex system, but an activated / identified particle sequence.

An N-level stratum thus produces particles that are the input stream for the higher level N + 1 layer. This self-consistent and fractal characteristic of the method according to the invention therefore makes it possible to analyze and segment the data stream at different levels of complexity.

It thus allows, always in the non-limiting example of the language to create more or less short text summaries.

It is interesting to note, with respect to the Pe ² / Cx threshold, that the numerator of this threshold represents the ergodic limit. The threshold thus calculated therefore measures the non-ergodicity of a layer within the meaning of the invention.

There is further provided a computer program product including instructions for performing the steps of the method described above when this program is executed by a processor. This program can be stored on a medium, for example a hard disk, a USB key (the "Universal Serial Bus"), a memory, downloaded via a telecommunications network, for example Internet, or other.

There is further provided a device for analyzing a complex system, comprising for example digital processing means, or else analog means, this device being arranged for: - provide an initial set of particle elements,

receive data forming a sequence of events produced by the complex system,

- determining for at least one, and preferably each, identified / activated particle element an index value from any detected strong interactions for that particle element, - selecting at least one particle element as a segmenter based on of said index value,

detecting events corresponding to segmenters in the received sequence and creating aggregates of particles, each aggregate corresponding to a sequence of events received in the sequence between two events corresponding to segmenters.

This device can for example include or be integrated in one or more processors, for example a microprocessor.

The invention will be better understood with reference to the figures, which illustrate embodiments given by way of example and not limiting.

Figure 1 schematically shows an example of particle element structure, according to one embodiment of the invention.

Figure 2 schematically illustrates an exemplary method according to an embodiment of the invention. We will now describe an embodiment of the invention implemented by a computer.

The method described below allows digital processing means to automatically and unconsciously correlate and segment a stream of digital data received by a computer, and to contextually organize the detected relationships and interactions.

The treatment may be independent of the nature of the data submitted.

This method implements an analysis system that can be related to a complex system in the sense of Poincaré. This system of analysis is non-specialized and with autosimilar topology (fractal), of interacting particles of particles.

This method can be applied to all domains for which the variability of the data is relatively complex or unknown, and of which it is difficult to make a reliable mathematical or statistical general modeling. For example, we can cite human written language, speech, meteorological models, molecular interaction models, DNA chains, and, in general, all numerical models of interacting systems of the massive interaction type.

For example, it is known from the prior art to detect a gene in an ACTG sequence of DNA from the messenger RNAs, from which a sequence of ACTG bases is deduced which is then sought in the DNA chain. This method, relatively long to implement, does not always ensure unequivocal recognition of the location of the desired gene because several identical forms may appear within the DNA chain examined.

The method illustrated in FIGS. 1 and 2 is based on a set of particle elements, or particles, in interactions according to relatively simple rules, topological and independent of the events submitted to the system. These rules make it possible to model the causality, in the quantum sense of the term, that is to say that there is no notion of explicit time, and that the notion of entanglement is used.

A particle within the meaning of this document (also called "event") is a fragment of the data stream received by the computer. For example, when the computer receives a stream of data consisting of a sequence of ASCII characters, a particle corresponds to a set of contiguous characters in that stream.

The computer represents each particle in its memory by a computer object. This computer object has a certain number of fields (segmentation index IS, identifier pID, excitation 1/0) and memory registers referenced 10 to 14 in FIG.

In the illustrated example, only one type of particles is provided, but the invention is not limited to this variant embodiment. We could alternatively provide several categories of particles.

The illustrated method thus implements a system of analysis that is complex, in the mathematical and physical sense of the term, without spatialization, and thus the structure adapts to that of the complex technical system that one seeks to analyze. Initially, a finite number of particles in a free state, non-discernible, non-materialized, is expected, although this number may be relatively large. Such a set could be called "contextual vacuum" by analogy with the quantum vacuum.

Indeed, when the computer receives the initial data flow, it does not know a priori how to segment this flow.

Data forming an event sequence produced by the complex system to be analyzed is received. When a new event is thus submitted to the analysis system, a particle is identified, or materialized, that is to say, one differentiates this particle to associate it with this event. In FIG. 2, the reference 21 designates a set of particles 22 materialized and associated with events received. Each materialized particle thus represents an event submitted to the analysis system.

Each materialized particle is identifiable, for example by a single number value. When an event corresponding to a materialized particle is received, this particle is activated.

With reference to FIG. 1, each materialized particle is associated with a segmentation index IS. This segmentation index is a counter that can be initialized, incremented and decremented. It represents at a given moment the number of interactions in which a particle participates.

Each materialized particle furthermore corresponds to a certain number of registers, including two registers for storing identifiers of particles assimilated to causes of this materialized particle: a so-called "strong causal coupler" register 11 for storing identifiers of the particles corresponding to the events preceding the event corresponding to this particle materialized in the received sequence of events, and

a so-called "contextual coupler" register 12 for storing an identifier of the particle preceding this materialized particle in a reordered particle stream during the execution of this analysis method, as explained hereinafter. Operation of register 11;

When the computer processes the data stream, it considers each fragment of this stream (in other words, each particle) and records in the register 11 of the computer object representing this particle, the identifiers of the other particles which immediately precede this particle in the flux. For example, in the case of an ASCII data stream, if the computer receives the string "AB", it records in the register 11 of the computer object "B" the pIDA identifier of the particle A. It is said that A and B in this case have a strong interaction because they follow each other in the stream. The computer increments the segmentation index IS of the computer object representing the particle A by one unit, to represent the fact that the particle A has a further strong interaction, in this case with the particle B

If it turns out that the same particle A preceded more than once a particle B in the stream (for example in the ABCAB stream), the pIDA identifier is even registered only once in the buffer 11 of B and the segmentation index of A is incremented only once.

In other words, the register 11 of the computer object representing a particle comprises a list of the identifiers of the particles which preceded at least once this particle in the stream and the segmentation index IS of a particle represents the number of different particles. which preceded this particle at least once in the flow.

In the embodiment described here, the buffer 11 of a particle is managed in a circular manner. Whenever the computer detects, for a particle B, a new occurrence of the same particle A which precedes B in the flow, it places the identifier of the particle A in the first rank in the register 11 of the particle B .

The ranks of the identifiers of two particles A, B in the register 11 of a particle C thus makes it possible to determine which of the two proximities AC, BC the computer has detected most recently.

Operation of register 12:

We have seen that the register 11 made it possible to represent the proximity interactions between the particles of the data stream, two particles having a proximity interaction if they follow each other in the stream. The invention uses another register 12 which operates in exactly the same way as the register 11 but instead of using as their distance between two particles their proximity in the stream, the difference between the segmentation indices of these particles is used as a distance. .

Seen from the computer, this amounts to considering a new particle stream in which the particles are classified according to their segmentation index

When the computer processes this new data stream, it again considers each fragment of this stream and records in the register 12 of the computer object representing this particle, the identifiers of the other particles that immediately precede this particle in this new stream. It is said that the particles have a weak interaction and the computer increments the segmentation index IS of this particle by one, to represent the fact that this particle has a weak interaction moreover.

The IS index of a particle is therefore the sum of its numbers of strong and weak interactions.

To summarize, two types of interactions are distinguished: the strong interaction, represented by the register 11, which concerns particles corresponding to events that are temporally close in the received sequence, namely immediately consecutive events. By analogy with particle physics it may be said that this is a short-range interaction;

- The weak interaction, the scope of which covers all the particles considered. This interaction is called internal segmentation interaction, or contextual interaction.

This contextual interaction is sensitive to the particle segmentation index, whereas the strong interaction is sensitive to causality in the received data stream. The analysis system implements similar rules for these two registers. Regarding the strong interaction, it is considered that each particle is in interaction with one or more particles preceding this particle in the received sequence. When the event corresponding to a first particle is received and is followed by an event corresponding to a second particle, this interaction has not been stored in memory, it is considered that the first particle has created an effect and its index of segmentation is incremented by 1.

At the second particle is associated the register 1 1, said strong causal coupler, to store identifiers of particles preceding this second particle (so-called effect particle). Each identifier is associated with a rank value within the register, the rank being a function of the age of the identification of the interaction with the effect particle associated with the register. When an interaction is detected, the particle preceding the effect particle sees its identifier brought to the register of the effect particle, with a rank equal to one, or, if this identifier already appeared in the register, its rank passes to one.

In addition, a particle identifier may be erased from the register if a lower-order particle identifier, already in the register, has its rank changed to one. The two necessary and sufficient cumulative conditions for the PIDB identifier of a particle B to be erased from a register 1 1 or 12 of a particle A is (i) that an identifier pIDc of a particle C of lower rank to that of B in this register sees its rank go to one and (ii) that the segmentation index IS of B is lower than the segmentation index IS of C. This is detailed below.

For example, if the sequence C, A, R, is received, the computer increments the segmentation index IS of the computer object representing the particle C, because C has a strong interaction with A. If the identifier corresponding to the event particle C initially did not appear in the register 1 1 of the event particle A, this identifier is brought to this register with a rank equal to one. The ranks of the possible identifiers already present in this register are incremented a (circular management of the register) or these identifiers are removed from this register, as explained below.

If the identifier corresponding to the event C initially appeared in the register 11 with a rank equal to RG strictly greater than one, the identifiers corresponding to particles of rank between one and (RG-1) have their rank incremented by one, and the others remain in their original ranks (circular permutation RG order), some identifiers being likely to be eliminated as explained below. If the identifier corresponding to the event C initially appeared in the register with a rank equal to one, the ranks of the identifiers of particles present in the register 11 are unchanged.

Concerning the register 1 1 associated with the event R, the identifier of the letter A is brought to this register with the rank one (it is supposed here that this identifier did not appear in the register of the event R). The segmentation index associated with event A is incremented by one.

If the sequence C, H, A, T is then received, immediately after the sequence C, A, R, the identifier of the letter C is brought to the register 11 corresponding to the letter H, always with the rank one, and the segmentation index of the letter C is incremented. Concerning the register associated with event A, the identifier of the letter H is entered in the register with a rank equal to one. The index associated with the event H is incremented by one.

As for event C, the identifier of this event is likely to be eliminated from the register of event A. More precisely, if the segmentation index IS corresponding to this event C is less than or equal to the index of segmentation of the event H having supplanted it in the register of the event A, then the identifier of the event C is eliminated from the register corresponding to the event A and the segmentation index of this event C is decremented. Otherwise, the identifier of the event C is kept in the register of the event A, with a rank equal to two. Among the rules implemented by the computer, one can also cite the following rule: if a given event corresponds to an identifier of higher rank RG to (RG-1) identifiers having interchanged their ranks (by circular permutation of order RG-1), the rank of this given event remaining a priori unchanged, and / or if this given event corresponds to an identifier supplanted in a given register by a new identifier (as in the example above), then we compare the segmentation index of this event given to the segmentation index of the new event (corresponding to the rank one identifier in the register). If the segmentation index of this given event is less than or equal to the segmentation index of the event corresponding to the rank one identifier in the register, then the identifier of this given event is removed from the register and the corresponding segmentation index decremented. Thus, each particle that sees one of its effects disappear has its segmentation index decremented by one. In other words, when an identifier is erased from the strong causal coupler of a particle, the segmentation index of this particle is decremented by one.

Conversely, when a new identifier is brought to the strong causal coupler of a particle, the segmentation index corresponding to this new identifier is incremented by one.

The segmentation index IS of a particle is thus a function of the number of causal interactions induced by the particle within the analysis system. As detailed below, this segmentation index is also a function of the weak (or contextual) interactions corresponding to the weak causal coupler register 12.

If a particle sees its segmentation index go down to zero, it returns to the free and undifferentiated state. The association of this particle with the corresponding event ends. In other words, this particle becomes indistinguishable.

The set of particles thus created by the computer following the reception of the computer stream of a sequence of events forms a contextual stratum, within which all the particles materialized are likely to interact. In FIG. 2, the reference 23o designates this stratum.

Automatic Segmentation of Computer Fllux It is recalled that an object of the invention is to automatically identify segmenters in the data stream. For this purpose, and as detailed below, the method according to the invention comprises a step of identifying segmenters in the data stream, the segmenters being the particles whose segmentation index IS is greater than a threshold value.

When a particle has a relatively high index of IS segmentation, that is to say that this particle is likely to interact (depending on the strong interaction and / or weak interaction) with a wide variety of particles, for example with all the other particles of the stratum, then this particle is likely to be considered as a segmenter. More precisely, the computer compares the segmentation index of this particle with a threshold, and if this index is greater than or equal to this threshold, said segmentation barrier, then the computer selects the particle as a segmenter. This threshold depends on the data stream received. More precisely, this threshold depends on the number of particles present in the stratum and the number of interactions detected.

For example, the value of the segmentation barrier, also referred to as the stratum segmentation index or the stratum IS, may be chosen equal to Pe ² / Cx, where Pe is the number of materialized particle elements of the stratum, and Cx is the number of strong and weak interactions detected between said materialized particles of the stratum.

In case of materialization or dematerialization of particle, addition or elimination of an identifier of a register, the value of the threshold is recalculated in order to take into account this change.

Particle aggregation The step of identifying the segmenters is followed by a step of aggregating the particles separated from each other by one or more consecutive segmenters.

As soon as at least one particle is selected as a segmenter, the computer aggregates between them the particles corresponding to events which in the incoming stream are separated from the other data of the sequence by two events corresponding to one or two segmenters. The particles aggregated together form a so-called composite particle. This composite particle is subject to the same rules as the other particles, that is to say that the computer creates a computer object for this particle and associates it in particular a segmentation index IS and a register 11 to memorize the strong interactions that she causes. In particular, a composite particle may be selected as a segmenter if its segmentation index is greater than the current stratum threshold.

The number Pe used to select the threshold value is a number of elementary particle elements, that is to say non-composite particles.

Thus, if the sequence C, I, R, A, G, E, space, D, E, B, O, T, T, E, S is received and the particles corresponding to the events "A, G, E" , "Space", "D, E" and "S" were selected as segmenters, so we aggregate them so as to obtain two additional particles, called composites:

- The particles corresponding to the events C, I, R, and - The particles corresponding to the events B, O, T, T, E.

Thus, this processing, based on segmenters determined from their segmentation index, makes it possible to identify incoming signal portions "cir" and "boot" having more meaning than the portions that would be obtained with a predetermined segmenter corresponding to the event "space", namely "waxing" and "débotté".

In addition, assuming that the particles corresponding to the events "E, R" and "O, N, S" were also selected as a segmenter, on the basis of, for example, the segmentation index generated by counting the interactions of which these events are the effects and / or the causes, then if the sequences C, I, R, O, N, S or C, I, R, E, R are received, the method allows to isolate the part of these words to which is attached the meaning of the word, that is to say "cir".

The method described above can thus make it possible to simplify the analysis of a complex system, since it can make it possible to overcome variations in the signal that finally become insignificant. For example, for a search engine implementing the method, the results obtained following the user request "waxing" may be the same as for the user query "wax".

More generally, the method described above makes it possible to discriminate the particles according to whether they are segmenters - we then speak of particlesΣ; or not - we are talking about nonΣ particles. If the nonΣ particles corresponding to consecutive events in the incoming flow and isolated from the other events of the flow by particlesΣ can be aggregated with each other, it is furthermore possible to aggregate particles corresponding to consecutive events in them. the incoming flow and isolated from the other events of the flow by nonΣ particles. For example, if the sequence C, I, R, A, G, E, space, D, E, space, B, O, T, T, E, S is received, it will be possible to create, in addition to the composite particles not Σ corresponding to the events C, I, R and B, O, T, T, E, a composite particle Σ corresponding to the events D, E, space. This composite particle can then play a role of segmentation.

The method may thus implement a nonΣ aggregator, referenced 24 in FIG. 2, and an aggregatorΣ, referenced 25 in FIG. 2, each of these aggregators comprising processing means and temporary storage means, for example a buffer. For example, the non agrég aggregator may be arranged to receive and store in the buffer memory identifiers of the particles not received, in an order corresponding to the order of arrival, and as long as no particle Σ n is received. When a particle is received, the contents of the buffer are copied to another memory, and then cleared from the buffer. When this other memory is thus written, a new composite particle is created. This composite particle is referenced 26 in FIG. 2. An aggregatorΣ can have a similar structure, except that it is the identifiers of the particlesΣ that are written in the buffer memory, and that it is the reception of a particle. nonΣ that triggers the creation of a new composite particle.

Thus, by detecting the strong interactions between particles in the incoming stream, one can dynamically estimate a segmentation index for each materialized particle. This segmentation index is used to detect particlesΣ and aggregate particles together. In addition, from these segmentation index values a flow of particles ordered by decreasing (or increasing in another embodiment) values of this segmentation index is created.

For each particle of this ordered stream, the register 12, called contextual coupler, is used to store the identifier of the particle preceding this particle, said particle effect, in this ordered stream. This storage obeys the same rules as register 11:

- In case of detection of a causal link between two consecutive particles in the ordered stream by segmentation index, the identifier of the previous particle in the ordered stream the effect particle is carried to the register of this particle effect with an equal rank to one (or sees its rank pass to one if that identifier was already present in the register), and its index of segmentation is incremented, If the identifier of the previous particle in the ordered flow the particle effect was already present in the register and that the transition to one of its rank has generated permutations of the ranks of identifiers of particles already present in the register and all having old and new ranks lower than that of the identifier of a third particle, then this third particle is likely to be eliminated from this weak interaction coupler. More specifically, if the segmentation index of this third particle is smaller (or less than or equal to) that of the previous particle in the ordered flow the particle effect then the identifier of the third particle is removed from the register and the index segmentation of the third particle is decremented.

The detection of the links in this ordered stream thus modifies the values of the segmentation indices. A new ordered flow is then developed based on the new segmentation indices, which makes it possible to detect new causal links, and thus to update the values of the segmentation indices, etc. again.

The values of the segmentation indices evolve dynamically, as a function of the strong interactions detected in the incoming flow and as a function of the weak interactions detected in the ordered flows.

When there is formation of aggregates, the additional particles thus created, like the particle referenced 26 in FIG. 2, also have a segmentation index and are thus also part of the ordered particle flows by decreasing values of these segmentation indices. .

Creating contexts

In this embodiment, the method according to the invention comprises a context creation step, two particles being in the same context if each of these particles comprises the identifier of the other particles in its register 12.

The registers 12 of the different particles also make it possible to detect contexts: if an identifier of a first particle appears in the context register 12 of a second particle and vice versa, then it is considered that these two particles form a causal loop. It will be advantageous to retain only the causal loops formed of at least two particles. A causal loop with for example three particles is detected if: a first particle is identified as the cause of a second particle, from the point of view of weak interaction, - this second particle is identified as the cause of a third particle, and

this third particle is identified as the cause of the first particle.

A causal loop may comprise more or fewer than three particles, for example two, four, five, six particles and more. In case of causal loop detection, a context is created. The method may thus implement a non-X contextualizer, referenced 27 in FIG. 2, and a contextualizer, referenced 28 in FIG. 2, each of these contextualizers comprising processing means and temporary storage means, for example a buffer. These contextualizers 27, 28 receive ordered flows of non-Σ and Σ particles respectively and include processing means arranged to detect the causal loops.

When a context is thus created following the detection of a causal loop based on the weak interaction, a new particle 29, corresponding to this context, is materialized in a higher order layer, referenced 231 in FIG. this stratum is likely to be created on this occasion if the current stratum was hitherto the highest order stratum.

This particle of the upper stratum is said to be entangled with the particles of the context of the current stratum. These particles of the current stratum are called the generators of the particle of the upper stratum.

Thus, a context created within a current stratum is treated as a new particle within the higher order stratum. This particle is subject to the rules stated above, that is to say that it is associated in particular with a segmentation index, a register for storing the identifiers of the other particles of the stratum. with which strong interactions will be detected, and that this particle is likely to be activated and abandoned.

Simply, a particle of a higher stratum is activated, or excited, when at least one, and preferably at least two, of its generators are activated. In addition, if the particle of the upper stratum is excited, all of its generators within the current stratum are also activated. In other words, the excitation of a particle of a higher order stratum can lead to excite other particles of this stratum. The particles of the higher order stratum are thus identified and activated, with a temporal order related to the received data. In other words, a new internal causal signal is created, constituted by the synchronized succession of materialized contextual particles.

If two particles of the upper stratum are activated one after the other, the segmentation index of one of the particles is incremented, and the identifier of the other particle is taken to the so-called strong causal coupler register of this particle.

Some particles in this upper stratum are likely to be selected as the segmentator of this stratum, again based on the segmentation indices of these particles and by comparing these indices to a stratum index established by applying the same formula as described. above.

The particles of this higher order stratum can be aggregated into composite particles, again based on the identification / activation sequence received at the input of the higher order stratum.

Also, it will be possible to rearrange the particles of this upper layer by decreasing segmentation index values, and to base detections of weak interactions on these ordered flows, which may lead to the detection of causal loops of at least three particles. of this higher order stratum. If a context is created within this higher order stratum, a new particle, corresponding to this context, is materialized in an even higher order stratum.

The method can thus be applied as many times as necessary, that is to say that in a stratum 23 _n no longer creates context. The process continues until the number of contextual strata is sufficient to model the causal flow of events entered in the zero-order stratum. During the analysis, the number of particles in the different strata, and the number of strata can increase or decrease dynamically.

During the execution of the method described above, the segmentation index of a particle belonging to a layer of order zero or higher is likely to go to zero: the particle is then dematerialized. Moreover, if it happens during the execution of this process that the last particle in the free state, among the initial number of particles, is materialized in a stratum, then we determine, starting from the stratum higher, what particles of the next lower stratum are the generators of these particles of this highest stratum. It goes back from stratum to stratum up to the zero order stratum receiving the sequence of events from the complex system. At each stratum, the particles that are not generators are dematerialized.

The particles are thus likely to be dematerialized if their index of segmentation goes to zero (we speak then of regulation with weak stress), but also if, the initial number of undifferentiated particles being exhausted, they did not lead to contexts significant (we speak then of regulation with strong constraint).

The weak constraint is always active, and consists, in a causal register (strong or weak), when a particle approaches (by its rank) of its particle effect, to break the coupling of the particle furthest from this particle effect , if the segmentation index of the latter is less than or equal to the segmentation index of this particle in approximation. The segmentation index of this most distant particle is then likely to decrease, and therefore to reach zero, which leads to the dematerialization of this particle. The strong constraint, called vacuum collapse, is achieved when the set initially planned no longer contains free particles. Couplings of particles that are not part of a causal loop are destroyed, and contexts are reorganized by eliminating the coupling of the particle with the lower segmentation index of each context. The process is initiated from the highest order contextual stratum.

Figure 1 shows an example of a computer object representing a materialized particle.

Each materialized particle, represented symbolically by the reference 1, is associated with a memory space for storing an identifier of the particle pID, a value of a boolean variable indicating the excitation state of the particle Excitation, this variable taking for example the value 1 when the particle is excited, and a segmentation index value IS of the particle.

In addition, are associated with the particle five registers:

Register 11 already described above, called strong causal coupler or CF,

A register 10, called a composite coupler or CC, in which the identifiers of the particles aggregated with the particle are stored, or alternatively identifiers of the composite particles incorporating the particle in question,

- The register 12, also called contextual coupler of the current stratum, that is to say the stratum to which the particle belongs, or ST0. The identifiers of the particles in the current stratum that form one or more contexts with the particle 1 are stored in this register.

A register 13, called lower contextual stratum coupler or ST-1, in which are stored identifiers of the generators of the particle 1, and A register 14 also known as a top-layer contextual coupler or ST + 1, in which are stored identifiers of the particle (s) of the upper layer having the particle 1 as a generator. In other embodiments, a memory space could also be provided for storing an identifier of the stratum to which the particle belongs, a binary inhibition, selection or other indicator.

Memory locations for each stratum created are also provided for storing a stratum identifier, the value of the segmentation threshold of that stratum, an identifier of the immediately lower stratum (parent stratum), and an identifier of the stratum immediately above ( child stratum), etc.

This method thus makes it possible to detect all significant forms of segmentation and relations between the data of any incoming signal, and to construct a context. This context itself becomes a datum transmitted to a higher stratum, within which meta-contexts can be constituted. These meta- contexts are in turn transmitted to an even higher stratum, etc. The process can be continued on a non-predetermined number of layers, that is to say as many times as necessary to model as completely as possible the relations and structures contained in the incoming signal.

In the case of written language, for example, the method described above can make it possible to identify the main segmenters of the language, then the possible structures constituting a signal of syntactic type, then the words, the phrases, as well as their contexts. membership. For example the word "black" can belong to:

- a common context with "neutron star", "astronomy", "black hole", etc.

- another context with "color", "red", etc., and

- another context with "skin", "civic rights", "Martin Luther King", etc. To resume the example of the gene detection described above, the knowledge of segmentation information can be used to discriminate the position of the gene to search more correctly than in the prior art.

Claims

claims

A method for analyzing a complex system, implemented by digital processing means, comprising: receiving a stream of digital data produced by the complex system

following receipt of a particle constituted by a fragment of said stream, creating and storing a computer object associated with this particle (22), said object comprising an identifier (pID); detecting strong interactions between said particles, each strong interaction corresponding to two consecutive particles in said stream,

storing in a so-called "strong causal coupler" register (11) of the computer object associated with each particle, the identifiers of the computer objects of the other particles having at least one strong interaction with said particle

determining for each particle a segmentation index value (IS) representative of the number of strong interactions detected for said particle and storing said index value (IS) in the computer object associated with said particle ",

selecting at least one particle as a segmenter of said stream when the value of the segmentation index (IS) associated with this particle exceeds a threshold value,

- creating composite particles consisting of aggregates of particles (26), each aggregate corresponding to a sequence of particles received in said stream between two particles corresponding to segmenters;

- Use said aggregates to prioritize and categorize the data of said flow.

2. Method according to claim 1, characterized in that it comprises a step of using said segmenters to minimize the memory required for encoding the data of said stream.

The method of claim 1 or 2, further comprising re-ordering said identified particles in an ordered sequence, based on the segmentation indices (IS) of said particles (22); detecting weak interactions between said particles, each weak interaction corresponding to two consecutive particles (22) in said ordered sequence, the segmentation index value (IS) stored in the computer object associated with a particle being representative of the number of strong and weak interactions detected for said particle, store in a so-called "contextual coupler" register (12) of the computer object associated with each particle, the identifiers of the computer objects of the other particles having at least a weak interaction with said particle create a context with at least two particles the infomatic object associated with each of said particles comprises in its contextual coupler register (12) the identifiers of the computer objects of each of the other particles; using said contexts to generate and store a compressed version of said data stream.

4. Method according to any one of claims 1 to 3, wherein the threshold is chosen equal to Pe ² / Cx, Pe being the number of particles (22) identified in said stream, and Cx being the number of weak interactions and / or strong detected between these particles.

The method according to any one of claims 1 to 4, wherein comparing the index value (IS) of each particle (22) with a low threshold, and wherein said particle is retained only if said value of index is above this low threshold.