Method for automatically generating a decision-making support algorithm; Product includes computer program and associated computer device.

Abstract

Method for automatically generating a decision-making support algorithm; Product includes computer program and associated computer device.This process (100) comprises the steps of: providing (110) logical rules, each associating a piece of knowledge with one or more physical data points; analyzing (120) the provided logical rules to extract a set of knowledge points and a set of physical data points; developing (130) a network of logical tensors comprising a neural network for each piece of knowledge point and a neural network for each physical data point, a neural network calculating the relevance of the knowledge point or associated physical data point from a current situation, defined by the current values ​​of the physical data points; training (140) the network on a training dataset; and executing (150) the decision-making support algorithm resulting from the training phase on a new situation to calculate the relevance of each piece of knowledge point and each physical data point with respect to this new situation. Figure for the abstract: Figure 1.

Description

Title of the invention: Method for automatically generating an algorithm to assist decision-making; Computer program product and associated computer device.

[0001] The present invention relates to the field of decision-making assistance for operators in complex and dynamic environments, in particular decision-making assistance for pilots of civil or military aircraft.

[0002] For decision-making in a complex and dynamic environment, the operator must, under strong time constraints, manage a large quantity of data describing the environment, only part of which is relevant for making a decision adapted to the particular situation encountered at the current time.

[0003] Making an appropriate decision then depends on the pilot's situational awareness of the environment.

[0004] Situation awareness can be defined as a three-level cognitive mechanism: (1) the perception of elements in the environment, (2) the understanding of their meaning, and finally (3) the projection of the future state of the significant elements.

[0005] This model highlights how the pilot captures environmental description data and analyzes it in order to make decisions. This model applies well to the case of pilots on commercial flights, a use case where this model has been the subject of numerous validations.

[0006] However, while he is focused on the mission to be carried out, the pilot may be led not to consider a piece of data because it is not significant for the mission to be carried out, whereas it is for the current situation. More generally, to deal with a situation, because of factors such as workload, stress, or simply distraction or fatigue, the pilot may ignore or forget information that would have allowed him to have better situational awareness.

[0007] It is clear that not taking into account certain environmental data in the first stages of the model can lead to erroneous awareness of the situation and therefore to making an inappropriate decision.

[0008] In certain operational phases, a transmitting operator, such as the pilot, is required to share environment description data with a receiving operator, such as the co-pilot (or the air traffic controller). These communications limit the amount of information to be shared. The transmitting operator does not share all the environment description data with the other operator, but chooses to share certain data that he knows constitutes a shared basis for common situational awareness. However, the transmitting operator can overload the information communicated by adding unnecessary information, due to a lack of filtering on its part of the available information.

[0009] There is therefore a problem linked to the operator's ability to master the information relevant to the particular situation encountered.

[0010] To ensure decision-making performance in a complex and dynamic environment, it is necessary to ensure the availability of relevant knowledge without overloading pilots with information that is not necessary.

[0011] It is necessary to reduce the quantity of information coming from the environment by describing and evaluating the relevance of this information through a decision support device.

[0012] However, the occurrence of a situation requires rapid decision-making. This is why decision support systems must optimize their calculations and provide the pilot in real time with the relevant information for understanding the situation and making decisions.

[0013] The aim of the present invention is to solve this problem.

[0014] For this purpose, the invention relates to a method for automatically generating an algorithm to assist in decision-making by an operator in a complex and dynamic environment, characterized in that the method comprises the steps of: providing a knowledge base comprising a plurality of logical rules, a logical rule being a logical function associating knowledge with one or more input variables, an input variable being physical data and / or intermediate knowledge; analyzing the knowledge base to extract a set of knowledge and a set of physical data;developing a logic tensor network type algorithm comprising a neural network for each piece of knowledge in the knowledge set and a neural network for each piece of physical data in the physical data set, a neural network calculating a relevance of the associated knowledge or physical data from a current situation, the current situation being defined by the current values ​​of each of the physical data in the physical data set; training the logic tensor network type algorithm on a learning database comprising situations annotated with the true relevance of each piece of knowledge and each piece of physical data, to obtain the decision-making assistance algorithm; and, executing the decision-making assistance algorithm on a new situation to calculate the relevance of each piece of knowledge and each piece of physical data with respect to the new situation.

[0015] According to particular embodiments, the method comprises one or more of the following characteristics, taken in isolation or in all technically possible combinations:

[0016] - analyzing the knowledge base allows to extract, for each knowledge of the knowledge set, a set of causes of said knowledge, which is made up of the physical data which are the direct or indirect causes of said knowledge, and in which the training of the algorithm of the logical tensor network type is carried out by implementing a constraint based on the set of causes of each knowledge.

[0017] - the training of the algorithm of the logical tensor network type is carried out in minimizing a global truth value resulting from an aggregation of a plurality of elementary truth values, the plurality of elementary truth values ​​comprises: at least the relevance of a physical datum for any situation s such that said physical datum is annotated as relevant; at least the relevance of a physical datum for any situation such that said physical datum is annotated as not relevant; and, for each knowledge, at least one difference between the relevance of said knowledge and the logical disjunction of the relevances of all the causes of said knowledge.

[0018] - the environment is that of an aircraft and the operator is the pilot of said aircraft.

[0019] - following the execution of the decision-making assistance algorithm on the new situation, the method comprises a step of presenting to the operator a list of knowledge and / or data filtered and / or ordered according to their respective relevance.

[0020] The invention also relates to a computer program product comprising software instructions which, when executed by a computer device, implement the decision-making assistance algorithm resulting from the implementation of the preceding method.

[0021] The invention also relates to a device configured so as to execute the decision-making assistance algorithm resulting from the implementation of the preceding method.

[0022] The invention and its advantages will be better understood on reading the detailed description which follows of a particular embodiment, given solely by way of non-limiting example, this description being made with reference to the appended drawings in which:

[0023] [Fig-1] [Fig.l] is a block representation of a mode of carrying out the identification method according to the invention; and,

[0024] [Fig.2] [Fig.2] is a schematic representation of the system for the implementation implementation of the identification method according to the invention.

[0025] [Fig.l] illustrates a preferred embodiment of a method 100 for developing an algorithm to assist decision-making by an operator in a complex and dynamic environment.

[0026] This method is carried out on a suitably programmed computer.

[0027] The developed algorithm is based on a neuro-symbolic architecture, combining the semantic aspect of symbolic artificial intelligence and the performance of neural networks.

[0028] In a step 110, a knowledge base is provided. In the example of the present embodiment, this is a weather knowledge base.

[0029] The knowledge base is, for example, a base prepared by experts, here meteorologists and aircraft pilots.

[0030] It can take different forms. For example, it is a text file. Alternatively, it is an ontology, that is to say a structured set of concepts and relationships between concepts, which offers a modeling of a field of knowledge, in this case the meteorological field.

[0031] The knowledge base contains logical rules. These are empirical rules defined by domain experts.

[0032] A rule is a logical function taking as input the current values ​​of a set of input variables and delivering as output the truth value of a piece of knowledge.

[0033] The set of input variables includes one or more physical data and / or one or more intermediate knowledge.

[0034] A physical datum d is a quantity whose instantaneous value is measured by a system, on board the aircraft or remotely.

[0035] The set S of physical data d measured makes it possible to describe the environment.

[0036] A situation s is constituted by the instantaneous values ​​of the physical data of the set S.

[0037] Knowledge k is an element of interpretation that allows for an abstract representation of the environment. Knowledge has a semantic dimension. Semantics allows for a meaning to be given to a sign (like a word) depending on the context.

[0038] The set K of knowledge k allows for a high-level objective representation of the environment. More precisely, the instantaneous truth values ​​of all knowledge k in the set K constitute an objective representation of the situation. This is in contrast to situation awareness, which concerns the interpretation that an operator can make of the situation, i.e. a subjective representation of the situation.

[0039] For example, the knowledge base stores a first set of rules relating to the knowledge “Icing”. This first set includes, for example, the rule:

[0040] “Temp_air < ​​0°C => Icing”

[0041] Thus, the fact that the value at the current instant of the physical data “air temperature”, Temp_air, is actually less than 0°C, leads to the knowledge of a risk of icing (that is to say that the current truth value of the knowledge “Icing” takes the value unity).

[0042] The knowledge base stores a second set of rules relating to the knowledge “Heat”. This second set includes, for example, the rule:

[0043] “Air_temp > 20°C => Heat”

[0044] Thus, the fact that the value at the current time of the physical data “air temperature”, Temp_air, is actually greater than 20°C, leads to the knowledge that it is hot (that is to say that the current truth value of the knowledge “Heat” takes the value unity).

[0045] The knowledge base stores a third set of rules relating to the “Wind” knowledge. This third set of rules includes, for example, the rules:

[0046] “Wind_speed > 30 kn => Wind”

[0047] Thus the fact that the value at the current instant of the physical data “wind speed”, Wind_speed, is actually greater than 30 kn, leads to the knowledge that there is wind (that is to say that the current truth value of the knowledge “Wind” takes the value unity).

[0048] “Wind_speed > 10 kn AND Wind_direction > 220°

[0049] AND Wind_direction < 320° => Wind »

[0050] Thus the fact that at the current instant the value of the physical data item “wind speed”, Wind_speed, is greater than 10 kn and that the value of the physical data item “wind direction”, Wind_direction, is between 220° and 320° (relative to the longitudinal direction of the aircraft) leads to the knowledge that there is wind (that is to say that the truth value of the knowledge item “Wind” takes the value unity).

[0051] “Wind_speed < 10 kn AND Wind_speed_gust < 10 kn => NOT Wind”

[0052] Thus the fact that the current value of the physical data item “wind speed”, Wind_speed, is less than 10 kn and that the current value of the physical data item “speed of wind gusts”, Wind_speed_gust, is less than 10 kn leads to the knowledge that there is no wind (that is to say that the truth value of the knowledge item “NOT Wind” takes the value unity, or that the truth value of the knowledge item “Wind” takes the value zero).

[0053] “Heat => Wind”

[0054] Thus, the knowledge that it is hot (the current truth value of the intermediate knowledge “Heat” taking the value unity) leads to the knowledge that it is windy (that is to say that the current truth value of the knowledge “Wind” takes the value unity).

[0055] In a step 120, the knowledge base is analyzed to determine the set K of knowledge, the set S of physical data describing the environment, and, advantageously, for each knowledge k of the set K, the set Sk of physical data (or causes) which impact the knowledge k, directly or indirectly through one or more intermediate knowledge.

[0056] The set Sk of causes of knowledge k is a subset of the set S.

[0057] For example, for the three previous sets of rules, the result of this analysis step is summarized by the following table:

[0058] [Tables 1] Knowledge Cause(s) Icing Air_temp Heat Air_temp Wind Wind_speed Wind_direction Wind_speed_gust Air_temp

[0059] Thus for this example, the set K of knowledge k is made up of the knowledge “Icing”, “Heat” and “Wind”.

[0060] Thus, the set S of physical data d is made up of the data “Wind_speed”, “Wind_direction”, “Wind_speed_gust” and “Temp_air”.

[0061] The set of causes associated with the knowledge “Icing” includes the physical data “Temp_air”, the set of causes associated with the knowledge “Heat” includes the physical data “Temp_air”, and the set of causes associated with the knowledge “Wind” includes the physical data “Wind_speed”, “Wind_direction”, “Wind_speed_gust” and “Temp_air”.

[0062] Thus the physical data of air temperature, Temp_air, is a cause of the knowledge “Icing”, but also of the knowledge “Heat” and of the knowledge “Wind”.

[0063] The following step 130 is an encoding step allowing the construction of the architecture of an LTN algorithm.

[0064] We note Pk(s) the relevance of knowledge k with respect to situation s (i.e. the current values ​​of the physical data of the set S).

[0065] Relevance is a variable between 0 and 1. When the relevance Pk(s) is close to unity, the current values ​​of the physical data of the set S are such that the relevance of knowledge k is acquired. When the relevance Pk(s) is close to zero, the current values ​​of the physical variables of the set S are such that the relevance of knowledge k is not acquired.

[0066] In the common usage case, the truth value of knowledge k and the relevance value of knowledge k are very similar, but this is not necessarily the case. For example, a truth value of 1 for the knowledge "Heat" is considered to denote that it is very hot and a truth value of 0 is considered to denote that it is very cold. The knowledge "Heat" is considered relevant when its truth value approaches one or the other of its extremes (0 or 1). The knowledge "Heat" is then very relevant (relevance value close to 1) when its truth value is close to 0.

[0067] Similarly, we note Pd(s) the relevance of the physical data d with respect to the situation s.

[0068] An LTN algorithm is an algorithm from the Logic Tensor Networks (LTN) family.

[0069] The LTN algorithm is implemented for example using the free software “Tensorflow” (registered trademark) and its interface “Python” (registered trademark).

[0070] Generally speaking, an LTN algorithm is based on logic derived from first-order logic and fuzzy logic. It allows a logical function to be converted into an object called a computational graph in the “Tensorflow” software. The function is decomposed into elementary predicates and the following principles are applied:

[0071] - constants and variables are represented by tensors;

[0072] - the elementary predicates associating these constants and these variables are represented by a function which may be a neural network; and,

[0073] - logical operators between elementary predicates are represented by fuzzy logic operators.

[0074] In the present method, the LTN algorithm is constructed so as to associate a neural network with each knowledge k of the set K and a neural network with each physical data d of the set S.

[0075] The neural network associated with knowledge k makes it possible to approximate the relevance Pk(s). This neural network therefore takes situation s as input and delivers as output the relevance of knowledge k with respect to situation s.

[0076] The neural network associated with the physical data d makes it possible to approximate the relevance Pj(s). It therefore takes as input the situation s and delivers as output the relevance of the data d with respect to the situation s.

[0077] The input of a neural network is a tensor whose coordinates are the current values ​​of the data in the set S.

[0078] In the example above, the set S being made up of the data “Wind_speed”, “Wind_direction”, “Wind_speed_gust” and “Temp_air”, the input tensor (30, 90, 30, 5) corresponds to the situation s in which:

[0079] - the wind speed is equal to 30 kn

[0080] - the wind direction is at an angle of 90°

[0081] - the speed of the gusts is 35 kn

[0082] - the air temperature is equal to 5°C.

[0083] In a step 140, the LTN algorithm is suitably parameterized by implementing a training phase.

[0084] The training uses batches of labeled situations, that is to say situations s annotated with the true value of the relevance of the different knowledge k of the set K and the true value of the relevance of the different data d of the set S. A training database is used, which includes the labeled situations. This database is for example developed from real situations annotated by experts.

[0085] The training makes it possible to adjust the parameters of the different neural networks of the LTN algorithm so that the relevances estimated by each of these neural networks are in line with the annotations.

[0086] Preferably, the training consists of maximizing a global truth value resulting from the aggregation of a plurality of elementary truth values. The aggregation is for example the product of the elementary truth values.

[0087] The plurality of elementary truth values ​​includes:

[0088] - the truth value equal to^^s) for any situation s such that the physical data d is annotated as relevant;

[0089] - the truth value equal to Not Pd{s) for any situation s such that the data physics d is annotated as not being relevant; and, advantageously; and,

[0090] - for each knowledge k of the set K, the truth value equal to 1-Pk(s) + Uz £StPd(s), which translates the axiom Pk(s) £SnPd(s), where "U" is the logical disjunction operator that is applied to all physical data d of the set Sk of causes of knowledge k (determined in step 120). This axiom translates the fact that the relevance of knowledge k for a situation s is equivalent to the combination of the relevances for situation s of each of the causes d of knowledge k.

[0091] Those skilled in the art will note that the contribution of logical rules in learning is made through the constraints formulated by these elementary truth values. Alternatively, other constraints (alternative to the preceding constraints or additional to the preceding constraints) could be developed from the relevance of the data and / or knowledge.

[0092] At each training step of the LTN algorithm, the global truth value is calculated for an annotated situation s. We preferably use a classic gradient descent with a cost function equal to unity minus the global truth value.

[0093] Among the set S of data d, it will mainly be the data d of the set Sk of causes of knowledge k which will influence the result of the calculation of the relevance of knowledge k, but it is possible that the training “discovers” new correlations between data and knowledge, that is to say correlations which were not present in the knowledge base provided initially. It is possible to enrich the initial knowledge base with new rules translating this new correlation.

[0094] The LTN algorithm, once trained, is none other than the decision-making assistance algorithm that can be used, in a step 150, to reason about new situations.

[0095] The algorithm is downloaded onto a decision-making support computer device on board the aircraft. This is illustrated by the device 200 of [Fig. 2], which is capable of executing the algorithm 230. The device is connected to sensors or systems (generally referenced by the number 210 in [Fig. 2]) allowing the acquisition of the instantaneous values ​​of the physical data constituting the inputs of the algorithm.

[0096] For a given situation, the decision-making assistance algorithm calculates the relevance of each knowledge k by executing the corresponding neural network and the relevance of each physical data d by executing the corresponding neural network.

[0097] Being able to obtain the relevance of each physical data allows for better explainability of the results on knowledge. Indeed, we can thus determine which physical data influenced the result, for example the relevance of knowledge k.

[0098] For example, by considering the knowledge "Wind" and its causes "Wind speed", "Wind direction", "Gust speed", and "air temperature" after training, we can obtain the relevance of each of these five pieces of information.

[0099] The relevance of each physical data makes it possible to determine which physical data have actually influenced the relevance of the wind knowledge.

[0100] For example:

[0101] - the relevance of the wind is high and the relevance of each of the four data physical is high, then each of the four data influenced the result in a similar way;

[0102] - the relevance of the wind is low and the relevance of each of the four data physical is weak, then each of the four data influenced the result in a similar way;

[0103] - the relevance of the wind is high and the relevance of one of the physical data is high, the others remaining low, then this physical data influenced the result more strongly than the others.

[0104] It is theoretically impossible for the relevance of a cause to be strong and that of knowledge weak. The use of the logical disjunction operator implies this property.

[0105] The outputs of the assistance algorithm can be used in different ways. For example, knowledge whose relevance is greater than a threshold is displayed on a screen 240 of the decision-making assistance computing device 200. In this way, the pilot is informed of the knowledge associated with the current situation. He therefore has an objective representation of the situation.

[0106] Alternatively, physical data whose relevance is greater than a threshold are displayed. The pilot's attention is then drawn to this information that is significant with respect to the current situation. It is then up to the pilot to develop his own representation of the situation, but from objectively filtered information. This makes it possible to lighten the pilot's cognitive load. This makes it possible to reduce the amount of information that the operator has to take into account by only displaying at each moment the information that is known to be useful (i.e. the physical data influencing the knowledge relevant to the situation encountered).

[0107] Alternatively, the physical data are ordered according to their current relevance so as to present to the pilot first the data which have the most influence on the understanding of the current situation or which are to be shared with the other members of the crew for the development of a common situation awareness.

[0108] This solution makes it possible to meet execution time constraints. It is therefore compatible with real time. Indeed, the architecture on which it is based has the following advantages:

[0109] - compared to deep learning systems, it requires less training data, while providing better generalization, interpretability and explainability of results.

[0110] - compared to symbolic artificial intelligence systems, it is of execution faster, with the ability to work with erroneous or incomplete knowledge bases and extract new knowledge.

[0111] In addition, after training, only the trained neural networks need to be embedded. The resources required are therefore low, both in terms of storage memory, computing capacity and execution time. This makes it possible to meet both embeddability and real-time constraints.

[0112] The proposed solution shows good results even with a reduced training base. Indeed, convincing results were obtained with 60 labeled situations in the training base and 30 training epochs, i.e. a training time of about fifteen seconds. Since developing a training base is a tedious task, it is a decisive advantage that the algorithm according to the invention can be trained efficiently with a small amount of data.

[0113] The proposed solution makes it possible to take into account business knowledge in automatic learning, through the analysis of an existing knowledge base.

[0114] Adapting the help algorithm to any modification of the initial knowledge base is easy. There is therefore a dynamic dimension to the present solution.

[0115] The symbolic aspect induces a semantic dimension allowing to give meaning to the results obtained. This explainability or interpretability of the results helps to certify the developed solution.

[0116] The invention can be used in environments other than that of civil or military avionics: control of autonomous cars, management of sensor services in airborne surveillance, distribution of resources, computing at the edge of the network, etc.

Claims

1. Claims A computer-implemented method (100) for automatically generating an algorithm (230) for assisting an operator in making a decision in a complex and dynamic environment, characterized in that the method comprises the steps of: - providing (110) a knowledge base comprising a plurality of logical rules, a logical rule being a logical function associating knowledge with one or more input variables, an input variable being physical data and / or intermediate knowledge, physical data being a quantity whose instantaneous value is measured by a system on board the aircraft or remotely, all of the measured physical data making it possible to describe the environment and knowledge being an element of interpretation making it possible to have an abstract representation of the environment, all of the knowledge allowing a high-level objective representation of the environment; - analyzing (120) the knowledge base to extract a set of knowledge and a set of physical data; - developing (130) an algorithm of the logical tensor network type comprising a neural network for each piece of knowledge in the set of knowledge and a neural network for each piece of physical data in the set of physical data, a neural network calculating a relevance of the knowledge or of the associated physical data from a current situation, the current situation being defined by the current values ​​of each of the physical data in the set of physical data; - training (140) the logic tensor network type algorithm on a learning database comprising situations annotated with the true relevance of each piece of knowledge and each physical data, to obtain the decision-making assistance algorithm; and, - execute (150) the decision-making assistance algorithm on a new situation to calculate the relevance of each piece of knowledge and each physical data with respect to the new situation.

2. Method according to claim 1, in which analyzing (120) the knowledge base makes it possible to extract, for each knowledge of the set of knowledge, a set of causes of said knowledge, which is made up of the physical data which are the direct or indirect causes of said knowledge, and in which the training of the algorithm of the logical tensor network type is carried out by implementing a constraint based on the set of causes of each knowledge.

3. Method according to claim 2, in which the training of the logic tensor network type algorithm makes it possible to adjust the parameters of the different neural networks of the logic tensor network type algorithm is carried out by minimizing a global truth value resulting from an aggregation of a plurality of elementary truth values, the plurality of elementary truth values ​​comprising: - at least the relevance of a physical data item, calculated by the associated neural network, for any situation such that said physical data item is annotated as relevant; - at least the relevance of a physical data item, calculated by the associated neural network, for any situation such that said physical data item is annotated as not relevant;and, - for each knowledge, at least one difference between the relevance of said knowledge, calculated by the associated neural network, and the logical disjunction of the relevances of all the causes of said knowledge, respectively calculated by the associated neural networks.;

4. A method according to any preceding claim, wherein the environment is that of an aircraft and the operator is the pilot of said aircraft.

5. Method according to any one of the preceding claims, in which, following the execution of the decision-making assistance algorithm on the new situation, the method comprises a step consisting of presenting to the operator a list of knowledge and / or data filtered and / or ordered according to their respective relevance.

6. Computer program product comprising software instructions which, when executed by a computing device, implement the decision-making assistance algorithm

7. decision resulting from the generation method according to any one of the preceding claims. A decision-making assistance computing device (200) configured to execute the decision-making assistance algorithm (230) resulting from the generation method according to any one of claims 1 to 5.

Images