Control method using recognition result of recognition module
The control method addresses the safety concerns of inaccurate recognition modules by using aggregation results and context information to estimate situation probabilities and optimize control commands, ensuring high safety probabilities and adaptable performance.
Patent Information
- Application Number
- PCT/KR2024/007066
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-26
- Filing Date
- 2024-05-24
- Publication Date
- 2024-12-05
AI Technical Summary
Existing control methods fail to guarantee safety with high probability despite the inaccuracy of recognition modules, especially in unforeseen situations.
A control method that uses the aggregation result from the aggregation module and context information to determine the actual situation, estimating the probability of each situation and determining an optimal control command to minimize the objective function while considering user-specified parameters and constraints, ensuring safety even in uncertain conditions.
Guarantees safety with a high probability by optimizing control commands based on recognition results, accounting for inaccuracy and unforeseen situations, and allows for learning optimal recognition module parameters to maximize control performance.
Smart Images

Figure KR2024007066_05122024_PF_FP_ABST
Abstract
Description
Control method using recognition results of recognition module
[0001] The present invention relates to a control method utilizing the recognition results of a recognition module, and more particularly, to a control method utilizing the recognition results of a recognition module capable of ensuring the safety of control with a high probability despite the inaccuracy of the recognition module, and capable of ensuring the safety of control even in situations not considered when creating the recognition module.
[0002] There are numerous studies that ensure control safety beyond a given probability. For example, K. Ren, H. Ahn, and M. Kamgarpour, “Chance-constrained trajectory planning with multimodal environmental uncertainty,” "IEEE Control Systems Letters, pp. 1-1, 2022." Furthermore, many recent studies have focused on prediction and optimization techniques, including Brandon Amos and J. Zico Kolter, “OptNet: Differentiable Optimization as a Layer in Neural Networks,” "OptNet: Differentiable Optimization as a Layer in Neural Networks, 2017."
[0003] Korean Patent Publication No. KR 10-2000-0073627 relates to a 'control system design method in a probability domain', and is a technology for improving the reliability of a control system by converting the governing equation in the physical domain into the probability domain so that a non-deterministic disturbance appearing as a function of time in the physical domain is converted into PSD (Power Spectral Density) appearing as a time-independent constant, and then controlling the system in the physical domain again with the control signal obtained.
[0004] Korean Patent Publication No. KR 10-2017-0005323 relates to a 'method for analyzing a system using a measurement signal', which calculates a deviation by subtracting a target value from the output of the system, accumulates the calculated deviation over time to calculate the net energy accumulated in the system, calculates an optimal cycle from the net energy, calculates a new net energy, and determines stability using the new net energy.
[0005] Therefore, the problem to be solved by the present invention is to provide a control method utilizing the recognition results of a recognition module that can guarantee the safety of control with a high probability despite the inaccuracy of the recognition module and can guarantee the safety of control even for situations that were not considered when creating the recognition module.
[0006] In addition, the present invention provides a computer-readable recording medium having recorded thereon a program for executing the above-described method on a computer.
[0007] The present invention provides a control method utilizing a recognition result of a recognition module, characterized in that, in order to achieve the above object, a control unit uses an aggregation result generated by an aggregation module and context information received from a context information providing unit to estimate a probability that an actual situation corresponds to an ith situation based on a recognition result in which a recognition module recognizes an external environment; and the control unit determines an optimal control command such that, for situations in which a control command does not satisfy a constraint condition, the sum of probabilities of the situations in which the control command does not satisfy the constraint condition is less than a maximum condition dissatisfaction probability, and the context information is a contextual probability of whether an actual situation corresponds to the ith situation.
[0008] In addition, the recognition module may further include a step of recognizing auxiliary data and generating a recognition result; and a step of the aggregation module aggregating the recognized auxiliary data and the recognition result to generate an aggregated result.
[0009] According to one embodiment of the present invention, the probability of the i-th situation can be estimated by further considering the significance level, which is the maximum allowable limit of the probability that the estimation range of the recognition module performance based on the aggregation result does not include the actual value due to the limited number of the auxiliary data.
[0010] The present invention provides a control method utilizing a recognition result of a recognition module, including a step in which a parameter and test result storage unit selects an artificial intelligence model parameter according to a user-specified parameter and transmits the user-specified parameter and the artificial intelligence model parameter to the recognition module; a step in which the recognition module recognizes a surrounding environment in consideration of the user-specified parameter and the artificial intelligence model parameter and transmits the recognition result to a control unit; and a step in which the control unit determines an optimal control command that minimizes an objective function in consideration of the recognition result, the user-specified parameter, and constraints.
[0011] According to another embodiment of the present invention, the method further includes: a step in which a recognition module receives user-specified parameters, and an approximate loss function calculation unit receives the user-specified parameters and the approximate parameters; a step in which the recognition module recognizes test data and transmits the test data and the recognition result to the approximate loss function calculation unit; and a step in which the approximate loss function calculation unit transmits a gradient value of the approximate loss function to the recognition module, wherein an artificial intelligence model parameter can be determined or improved based on the gradient value transmitted to the recognition module.
[0012] The present invention provides a control method utilizing a recognition result of a recognition module, including a step in which a parameter and test result storage unit selects suitable artificial intelligence model parameters according to context information, user-specified parameters, significance levels, and maximum conditional dissatisfaction probability, and transmits the artificial intelligence model parameters to the recognition module; a step in which the recognition module recognizes an external environment according to the context information, the user-specified parameters, the significance levels, and the maximum conditional dissatisfaction probability, and transmits the external environment recognition result to a control unit; a step in which the control unit estimates the maximum probability of each situation in consideration of the context information, the user-specified parameters, the external environment recognition result, and the significance levels; a step in which a constraint is determined based on the estimated maximum probability of each situation and the maximum conditional dissatisfaction probability; and a step in which the control unit determines an optimal control command that minimizes an objective function in consideration of the external environment recognition result, the user-specified parameters, and the constraints.
[0013] According to another embodiment of the present invention, the method further includes: a step in which a recognition module receives a user-specified parameter, a learning conservative parameter, and a maximum conditional dissatisfaction probability, and an approximate loss function calculation unit receives the user-specified parameter, the maximum conditional dissatisfaction probability, and the approximate parameter; a step in which the recognition module and the approximate loss function calculation unit receive contextual information; a step in which the recognition module performs recognition on test data and auxiliary data and transmits a recognition result to the approximate loss function calculation unit; and a step in which the approximate loss function calculation unit transmits a gradient value of the approximate loss function to the recognition module, wherein the artificial intelligence model parameters can be determined or improved based on the gradient value transmitted to the recognition module.
[0014] In order to solve the above other technical problems, the present invention provides a computer-readable recording medium having recorded thereon a program for executing a control method utilizing the recognition result of the above-described recognition module on a computer.
[0015] According to the present invention, the safety of control can be guaranteed with a high probability despite the inaccuracy of the recognition module, and the safety of control can be guaranteed even for situations that were not considered when creating the recognition module.
[0016] In addition, according to the present invention, it is possible to learn optimal recognition module model parameters according to the user's preference, and control can be performed to ensure safety by utilizing recognition results for contextual information and auxiliary data even under uncertain situations.
[0017] Furthermore, according to the present invention, it is possible to learn optimal recognition module model parameters to maximize control performance while ensuring safety under specific contextual information and user preferences.
[0018] Figure 1 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred first embodiment of the present invention.
[0019] FIG. 2 is a flowchart of a control method utilizing the recognition result of a recognition module according to a preferred first embodiment of the present invention.
[0020] Figure 3 is a flowchart of the process in which the test unit (130) tests the recognition module (140).
[0021] Figure 4 is a flowchart of a general control among control methods utilizing the recognition results of a recognition module according to a preferred first embodiment of the present invention.
[0022] Figure 5 is a flowchart of robust control among control methods utilizing recognition results of a recognition module according to a preferred first embodiment of the present invention.
[0023] Figure 6 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred second embodiment of the present invention.
[0024] Figure 7 is a flowchart of a learning method for performing control using the recognition results of a recognition module according to a preferred second embodiment of the present invention.
[0025] Figure 8 is a flowchart of a control method utilizing the recognition result of a recognition module according to a preferred second embodiment of the present invention.
[0026] FIG. 9 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred third embodiment of the present invention.
[0027] Figure 10 is a flowchart of the learning process of the recognition module when safety is ensured using auxiliary data according to a preferred third embodiment of the present invention.
[0028] Figure 11 is a flowchart of a control method for ensuring safety using auxiliary data according to a preferred third embodiment of the present invention.
[0029] The present invention relates to a control method utilizing a recognition result of a recognition module, and comprises a step of a control unit estimating a probability that an actual situation corresponds to an ith situation based on a recognition result in which the recognition module recognizes an external environment by using an aggregation result generated by an aggregation module and context information received from a context information provider, and a step of the control unit determining an optimal control command such that the sum of probabilities of situations in which a control command does not satisfy a constraint condition is less than a maximum condition dissatisfaction probability, wherein the context information is a contextual probability of whether an actual reality corresponds to the ith situation, and the control method can ensure the safety of control with a high probability despite the inaccuracy of the recognition module, and can ensure the safety of control even for situations that were not considered when creating the recognition module.
[0030] Hereinafter, with reference to the attached drawings, preferred embodiments of the present invention will be described in detail, enabling those skilled in the art to easily implement the invention. However, these embodiments are intended to illustrate the invention more specifically, and it will be apparent to those skilled in the art that the scope of the present invention is not limited thereby.
[0031] In order to clearly solve the problem to be solved by the present invention, the composition of the invention will be described in detail based on a preferred embodiment of the present invention with reference to the accompanying drawings. In assigning reference numbers to components in the drawings, the same reference numbers are assigned to the same components even if they are in different drawings, and it is made clear in advance that components in other drawings may be cited when necessary when describing the drawings. In addition, in explaining the operating principle of the preferred embodiment of the present invention in detail, if it is determined that a specific description of a known function or configuration related to the present invention and other matters may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.
[0032] Additionally, throughout the specification, when a part is said to be 'connected' to another part, this includes not only cases where it is 'directly connected' but also cases where it is 'indirectly connected' with another element in between. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, "comprises" or "comprising" does not exclude the presence or addition of one or more other elements, steps, operations, or elements other than the mentioned elements, steps, operations, or elements.
[0033]
[0034] Figure 1 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred first embodiment of the present invention.
[0035] Referring to FIG. 1, a control device utilizing the recognition result of a recognition module according to the present embodiment is composed of a first data supply unit (110), a test unit (130), a recognition module (140), an aggregation module (150), a context information provision unit (160), and a control unit (180).
[0036] In a preferred first embodiment of the present invention, the recognition module (140) is considered as being provided externally, and the performance of the recognition model (the probability of each recognition result for each actual situation) is evaluated, and then the performance is controlled to ensure safety using the evaluation.
[0037] The first data supply unit (110) provides auxiliary data for recognition related to essential conditions that must be satisfied despite the inaccuracy of the recognition module (140), for example, to ensure safety.
[0038] The test unit (130) tests the recognition module (140) in a situation identical to or similar to an actual control situation based on auxiliary data supplied from the first data supply unit (110).
[0039] Auxiliary data may be voice, video, natural language, etc., depending on the nature of the recognition module (140).
[0040] The recognition module (140) recognizes external situations and generates electronic information about the external situations.
[0041] The recognition module (140) is a device that recognizes, receives, and processes non-standardized information such as sound and image, including sensors, their post-processing devices, and natural language processing modules. If the output of the recognition module (140) is continuous, it is preferable to discretize it and treat it as a discrete variable.
[0042] The aggregation module (150) receives and aggregates the recognition results of the recognition module (140) for each auxiliary data. It is preferable that the aggregation module (150) generates an aggregation table by matching each auxiliary data with the recognition results of each auxiliary data.
[0043] The context information provision unit (160) transmits context information to the control unit (180). Context information refers to the probability that each situation will occur under a specific environment.
[0044] For example, when recognizing a speed limit sign in a highway environment, the probability that the speed limit is 80 km / h or less is greater than the probability that the speed limit is 30 km / h or less, and contextual information refers to the probability of each situation under these specific environments.
[0045] The control unit (180) controls the target device (190) based on the external environment recognition result of the recognition module (140), the aggregation result of the aggregation module (150), and the context information provided by the context information provision unit (160).
[0046] The roles of the control unit (180), recognition module (140), and contextual information provider (160) can be performed by the same device. Additionally, depending on the situation, the roles of the test unit (130), control unit (180), recognition module (140), and aggregation module (150) can be performed by the same device.
[0047]
[0048] FIG. 2 is a flowchart of a control method utilizing the recognition result of a recognition module according to a preferred first embodiment of the present invention.
[0049]
[0050] At step 200, all inaccurate variables are eliminated from the system dynamics model of the controlled device and its surrounding environment. It is desirable to eliminate these variables using their initial values and control commands.
[0051] All inaccurate initial values and system dynamic model parameters in the entire system dynamic model are considered as recognition targets of the recognition module (140). Since all inaccurate initial values and system dynamic model parameters do not change, they can be considered as recognition targets, and when the elimination of inaccurate variables in the system dynamic model is completed, only variables whose values are known accurately and control commands remain in the dynamic model, except for all inaccurate initial values and system dynamic model parameters that are recognition targets of the recognition module (140) at the time of control start.
[0052] In step 210, a set of constraints is established by collecting the constraints that the variables in the system dynamics model must satisfy in each situation.
[0053] A constraint set refers to conditions that impose constraints on control when utilizing the recognition results of the recognition module according to the first embodiment of the present invention. For example, constraints may include a speed limit of 100 km / h on highways or a speed limit of 30 km / h in school zones.
[0054] At step 220, it is determined whether the test results of the test unit (130) require an update. It is desirable to conduct the test of the test unit (130) on the recognition module (140) initially and when testing is necessary (when the situation has changed or a significant amount of time has passed). The user or the control unit (180) can determine whether to re-conduct the test, taking into account the degree of change in the situation and legal safety standards.
[0055] If it is determined at step 220 that an update of the test results is necessary, proceed to step 230. If it is determined that an update of the test results is unnecessary, proceed to step 240.
[0056] At step 230, the recognition module (140) is retested based on the auxiliary data supplied from the first data supply unit (110), and the test results are updated in the aggregation module (150). The process of updating the aggregation module (150) is illustrated in FIG. 3.
[0057] At step 240, the control unit (180) determines whether to perform normal control or robust control. If normal control is performed, the process proceeds to step 250, and if robust control is performed, the process proceeds to step 260.
[0058] Robust control is an additional level of significance considered in addition to general control, and the user can decide whether to use robust control.
[0059]
[0060] Figure 3 is a flowchart of the process in which the test unit (130) tests the recognition module (140).
[0061] At step 300, auxiliary data is input from the first data supply unit (110) to the test unit (130). At this time, it is preferable that the auxiliary data be data on recognition related to essential conditions that must be satisfied despite the inaccuracy of the recognition module (140), for example, such essential conditions to ensure safety.
[0062] At step 310, the recognition module (140) is tested with auxiliary data.
[0063] At step 320, auxiliary data and test results are provided to the aggregation module (150).
[0064] It is desirable for the test unit (130) to test the recognition module (140) with auxiliary data while simultaneously providing the auxiliary data to the aggregation module (150).
[0065] At step 330, the recognition result of the recognition module (140) is provided to the aggregation module (150), and the aggregation module (150) aggregates the recognition result for each auxiliary data.
[0066] At step 340, the aggregation module (150) updates the aggregation result and provides it to the control unit (180).
[0067]
[0068] Figure 4 is a flowchart of a general control among control methods utilizing the recognition results of a recognition module according to a preferred first embodiment of the present invention.
[0069] At step 400, the control unit (180) receives context information from the context government providing unit (160).
[0070] At step 410, the control unit (180) receives the maximum conditional dissatisfaction probability from the user.
[0071] At step 420, the recognition module (140) recognizes the external environment and transmits the recognition result to the control unit (180).
[0072] At step 430, the control unit (180) estimates the probability of each situation (the probability that the recognized environment is each situation) using Bayes' theorem as shown in the following mathematical expression 1. That is, considering the recognition result recognized by the recognition module (140), the probability that the actual situation corresponds to the ith situation is estimated using the context information, the aggregation result of the aggregation module (150), and the recognition result.
[0073]
[0074] In mathematical expression 1 is the reality of the i-th situation, is the output of the recognition module (140). It means that. and is the output of the recognition module (140). The probability that reality is the i-th situation, is the output The probability of one day, When the reality is the i-th situation, the recognition module The probability of outputting , and is the number of test data in the i-th situation according to the test aggregation results, and among them, the result It refers to the number of test data printed.
[0075] is a contextual probability that the actual reality is the i-th situation, and is a value received from the context information provider (160) as context information.
[0076]
[0077] At step 440, the control unit (180) determines and executes a control command according to the following mathematical expression 2 as a solution to the optimization problem. It is preferable that the control unit (180) determines an optimal control command such that the sum of the probabilities corresponding to situations in which the constraints are not satisfied is less than the maximum condition dissatisfaction probability for situations in which the constraints are not satisfied.
[0078]
[0079] In mathematical expression 2, k and K are the current step and the final step, respectively, x and u are the state variable and the control command, respectively, J is the objective function of the control, and x k+1 =f k (x k ,u k ) is a system dynamics model.
[0080] C i is a set of state variables that must be included to satisfy the conditions of the constraint set for the i-th situation. When a state variable is included in this set, all conditions of the constraint set for the corresponding situation are satisfied. ε is the maximum probability of condition dissatisfaction, which is preferably given by the user. Here, the probability that situation i is real is p i is obtained from mathematical formula 1. Use the values for the actual output.
[0081]
[0082] Figure 5 is a flowchart of robust control among control methods utilizing recognition results of a recognition module according to a preferred first embodiment of the present invention.
[0083] At step 500, the control unit (180) receives context information from the context government providing unit (160).
[0084] At step 510, the control unit (180) receives a significance level r from the user. α and receives the maximum probability of conditional dissatisfaction. The significance level means the maximum allowable limit of the probability that the estimated range of the recognition module performance based on the aggregated results does not include the actual value due to the limited number of auxiliary data.
[0085] At step 520, the recognition module (140) recognizes the external environment and transmits the recognition result to the control unit (180).
[0086] At step 530, the control unit (180) recognizes that when the reality is the i-th situation, the recognition module is as in the following mathematical expression 3. The probability of outputting Calculate the range of .
[0087]
[0088] Here is the significance level Indicates the standard normal distribution value in , Is It can be chosen as a collection of arbitrary positive numbers satisfying .
[0089] The output of the recognition module (140) under the given significance level at step 540 The probability that reality is the i-th situation The maximum value can be calculated as shown in the following mathematical formula 4.
[0090]
[0091] In the robust control at step 550, the maximum value of Equation 4 is Use instead (p i) solves the optimization problem of mathematical expression 2 and determines and executes the control command according to the result. It is preferable that the control unit (180) determines the optimal control command such that the sum of the probabilities corresponding to the situations in which the control command does not satisfy the constraint condition is less than the maximum condition dissatisfaction probability for situations in which the control command does not satisfy the constraint condition. Through the robust control according to Fig. 5, safety can be guaranteed by considering the error according to the number of test data.
[0092] Figure 6 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred second embodiment of the present invention.
[0093] Referring to FIG. 6, a control device utilizing the recognition result of a recognition module according to the present embodiment is composed of a second data supply unit (120), a test unit (130), a recognition module (140), an approximate loss function calculation unit (155), a parameter and test result storage unit (165), and a control unit (180).
[0094] In a second preferred embodiment of the present invention, when performing control based on the results of a recognition module (140) using a deep learning method, a prediction and optimization (predict+optimize) method for training a recognition module (140) in a customized manner to maximize control performance is extended and improved using an approximate loss function so that it can be applied to general optimization problems and dynamic models and reflect user preferences.
[0095] The second data supply unit (120) provides test data for the control environment. The test data may be voice, video, natural language, etc., depending on the nature of the recognition module.
[0096] The test unit (130) is a device that can test the recognition module (140) in a situation identical to or similar to an actual control situation based on supplied test data.
[0097] The recognition module (140) according to a preferred second embodiment of the present invention preferably includes a deep neural network. In the preferred second embodiment of the present invention, the recognition module (140) preferably receives user-specified parameters and performs recognition accordingly to facilitate control considering user-specified parameters that reflect the user's preferences rather than simply recognizing the external environment, through control that takes user-specified parameters into account.
[0098] The system operator can transmit user-specified parameters to the recognition module (140) on behalf of the user during the recognition module learning process through a separate terminal, and transmit user-specified parameters and approximation parameters to the approximation loss function calculation unit (155).
[0099] The approximate loss function calculation unit (155) receives the test results from the test unit (130), calculates the approximate loss function in a manner to be described later under the user-specified parameters and approximate parameters given by the operator, obtains the gradient thereof, and provides the approximate loss function and the gradient thereof for learning the artificial intelligence model parameters of the recognition module (140).
[0100] The parameter and test result storage unit (165) stores the parameters of the artificial intelligence model within the recognition module (140) for which learning has been completed and the aggregated test results corresponding to the parameters.
[0101] The parameter and test result storage unit (165) transmits artificial intelligence model parameters suitable for user-specified parameters to the recognition module (140) together with the user-specified parameters.
[0102] The control unit (180) receives user-specified parameters according to the user's preferences.
[0103] The control unit (180) transmits the input user-specified parameters to the parameter and test result storage unit (165), and when the recognition module (140) recognizes the external environment, it receives the recognition result from the recognition module (140), solves an optimization problem according to the recognition result under the user-specified parameters, calculates the optimal control command, and transmits it to the controlled device (190).
[0104] The roles of the second data supply unit (120), test unit (130), recognition module (140), approximate loss function calculation unit (155), and parameter and test result storage unit (165) can be performed within the same device (central server, etc.).
[0105]
[0106] Figure 7 is a flowchart of a learning method for performing control using the recognition results of a recognition module according to a preferred second embodiment of the present invention.
[0107] Hereinafter, the system design and initialization process required for performing control according to a preferred second embodiment of the present invention will be described. First, all inaccurate variables are eliminated from the system dynamic model. At this time, the initial values and control commands of the corresponding variables are used. All inaccurate initial values and model parameters throughout the model are considered as recognition targets of the recognition module. Since the initial values and model parameters do not change, when the elimination is completed, only the variables and control commands whose values are known accurately remain in the dynamic model, other than the parameters and initial values that are recognition targets of the recognition module at the time of control initiation. Using the dynamic model thus constructed, an optimization problem for determining the optimal control command based on the recognition results is constructed as shown in the following mathematical equation 5.
[0108]
[0109] Here, x and z are continuous and discrete control variables, respectively, o is the recognition result of the recognition module, and r is a user-specified parameter. This optimization problem has an objective function constraints By minimizing under, and is determined by the system designer, taking into account the dynamic model and required control specifications.
[0110] Finally, the loss function L(x,z,θ,r), which is an evaluation of the control command of this control system, is defined. Since the loss function is a post-evaluation of the control command rather than a judgment during the control process, the actual environment θ is used instead of o, which is the recognition result of the recognition module (140).
[0111] The control method utilizing the recognition results of the recognition module according to the second embodiment of the present invention ultimately operates with the goal of minimizing a loss function. The process up to this point represents the first step in implementing the invention and may be an initial system design stage that is not included in the model learning or control process.
[0112]
[0113] In a learning method for performing control using the recognition result of the recognition module according to a preferred second embodiment of the present invention, in step 700, an operator of the model learning process transmits a user-specified parameter to the recognition module (140) and transmits the user-specified parameter and the approximation parameters λ and β to the approximation loss function calculation unit (155). (λ is a scalar and β is the number of constraints in the optimization problem) is a vector with the dimension of .)
[0114]
[0115] At step 710, test data is input from the second data supply unit (120) to the test unit (130).
[0116] At step 720, the recognition module (140) recognizes the input test data. The test unit (130) transmits the test data (meaning accurate information (θ) about the environment) to the approximate loss function calculation unit (155).
[0117] When the test is completed at step 730, the recognition module (140) transmits the recognition result to the approximate loss function calculation unit (155), and the approximate loss function calculation unit (155) calculates the approximate loss function as follows.
[0118]
[0119] At this time If U is a compact set for x when all are continuous and U contains all solutions of Equation 5 and U is restricted for each z, then the user-specified parameter λ is sufficiently small and each element of the approximation parameter β is sufficiently large. It converges pointwise to the L value for the solution that minimizes L among the optimal solutions of the optimization problem.
[0120] Therefore, in the second embodiment of the present invention, an approximate loss function The recognition module (140) is trained using L instead of the loss function for the control command (i.e., not a function of the recognition result of the recognition module). Now, for training, The (approximate) slope is found as follows.
[0121] i) Approximate gradient for discrete elements of recognition result o: In general, it is not possible to obtain the gradient for discrete elements, but in most cases, the discrete elements of recognition result are obtained as a function of some continuous variable (argmin, rounding, etc.). Discrete element o di is a vector w consisting of some continuous variables i (w i is an intermediate result within the recognition module.) The function f(w i ), f(w i ) is a suitable softened probability distribution p(o di :w i ) can be replaced with. As an example, for μ that follows a standard normal distribution, o di = f(w i +μ) can be used. If this softening is needed, f(wi ) settings This should be done in the early system design stage, not when finding the (approximate) slope of the softened probability distribution p(o di :w i ) is given as a result of the system design phase, the expected approximation loss function and its approximate gradient are obtained as follows.
[0122]
[0123] O here di is o di It means a set of possible values, and in the present invention, O di We only deal with the case where all of them are finite sets.
[0124] ii) Approximate gradient for a continuous element of recognition result o: one continuous element o ci The approximate slope for is sufficiently small ρ (if ρ is required, how to set ρ or its value is (which is determined at the initial system design stage, not when finding the (approximate) slope) can be defined as follows.
[0125]
[0126] iii) The (approximate) gradient of the special case where the following values are defined for the continuous elements of the recognition result o:
[0127]
[0128] This value can be used as an approximate slope. In some cases, it may be the exact slope. Here
[0129]
[0130] am.
[0131] Approximate loss function at step 740 The gradient value of the received approximate loss function is returned to the recognition module (140) (backpropagation begins). It is preferable that the recognition module (140) use the gradient value of the received approximate loss function to set or improve the artificial intelligence model parameters.
[0132] When the conditions specified by the operator are initially satisfied, learning is stopped, and the recognition module (140) sets the model parameters, and the loss function uses the aggregated test results (each , for ) is transmitted to the parameter and test result storage unit (165) and stored. If the condition is not satisfied, the process returns to step 710 and repeats this process.
[0133]
[0134] Figure 8 is a flowchart of a control method utilizing the recognition result of a recognition module according to a preferred second embodiment of the present invention.
[0135] At step 800, the control unit (180) receives user-specified parameters from the user.
[0136] In step 810, user-specified parameters are passed to the parameter and test result storage unit (165).
[0137] At step 820, the parameter and test result storage unit (165) selects model parameters suitable for the artificial intelligence model according to the user-specified parameters and transmits them to the recognition module (140) together with the user-specified parameters.
[0138] At step 830, the recognition module (140) recognizes the environment and transmits the recognition result to the control unit (180).
[0139] At step 840, the control unit (180) solves the optimization problem of mathematical expression 5 that minimizes the objective function by considering user-specified parameters and constraints based on the recognition result, determines the control command according to the optimal solution, and controls the controlled device (190).
[0140]
[0141] FIG. 9 is a configuration diagram of a control device utilizing the recognition result of a recognition module according to a preferred third embodiment of the present invention.
[0142] Referring to FIG. 9, a control device utilizing the recognition result of a recognition module according to the present embodiment is composed of a first data supply unit (110), a second data supply unit (120), a test unit (130), a recognition module (140), an approximate loss function calculation unit (155), a context information provision unit (160), a parameter and test result storage unit (165), and a control unit (180).
[0143] In a third preferred embodiment of the present invention, the first and second preferred embodiments of the present invention are combined to create a prediction and optimization-based control system that reflects user preferences and can be applied to a wide range of optimization problems, while also ensuring safety by utilizing auxiliary data supplied by the first data supply unit (110).
[0144] That is, in a preferred third embodiment of the present invention, safety is ensured using auxiliary data, and recognition module training is also performed.
[0145] The first data supply unit (110) provides auxiliary data for recognition related to essential conditions that must be satisfied despite the inaccuracy of the recognition module (140), for example, conditions essential to ensure safety.
[0146] The second data supply unit (120) provides test data for the control environment. The test data may be voice, video, natural language, etc., depending on the nature of the recognition module.
[0147] The test unit (130) can test the recognition module (140) in a situation identical or similar to an actual control situation based on the supplied test data.
[0148] During the recognition module learning process, the operator transmits user-specified parameters, learning conservative parameters, and maximum conditional dissatisfaction probability to the recognition module (140) on behalf of the user, and transmits user-specified parameters, learning conservative parameters, maximum conditional dissatisfaction probability, and approximation parameters to the approximation loss function calculation unit (155).
[0149] The recognition module (140) is a form capable of learning, such as a deep neural network, and receives user-specified parameters, learning conservative parameters, context information, and maximum condition dissatisfaction probability and performs recognition accordingly.
[0150] The recognition module (140) according to the third preferred embodiment of the present invention can be tested for accuracy with auxiliary data, and test data can be supplied to the recognition module (140) from the outside to be learned through a deep neural network.
[0151] The approximate loss function calculation unit (155) receives test results for auxiliary data of the first data supply unit (110) and test data of the second data supply unit (120), aggregates the results for the auxiliary data, and then calculates an approximate loss function under user-specified parameters, learning conservative parameters, maximum condition dissatisfaction probability, and approximate parameters given by the operator and context information provided by the context information providing unit (160), obtains a gradient, and provides an approximate loss function and its gradient for learning artificial intelligence model parameters of the recognition module (140).
[0152] The context information provision unit (160) provides context information to the recognition module (140) and the approximate loss function calculation unit (155). Context information refers to the probability that each situation will occur under a specific environment.
[0153] The parameter and test result storage unit (165) stores the artificial intelligence model parameters of the recognition module for which learning has been completed and the aggregated test results corresponding to the parameters.
[0154] The roles of the recognition module (140), the first data supply unit (110), the second data supply unit (120), the test unit (130), the approximate loss function calculation unit (155), the context information provision unit (160), and the parameter and test result storage unit (165) can be performed by the same device (central server, etc.).
[0155] Meanwhile, when safety is ensured using auxiliary data, the operation of each component during control is as follows.
[0156] The context information providing unit (160) provides context information to the control unit (180).
[0157] The user inputs user-defined parameters, significance level, and maximum condition dissatisfaction probability according to his / her preference into the control unit (180).
[0158] The parameter and test result storage unit (165) selects artificial intelligence model parameters according to user-specified parameters, significance levels, maximum conditional dissatisfaction probability, and contextual information, and transmits the artificial intelligence model parameters to the recognition module (140) together with the user-specified parameters, significance levels, maximum conditional dissatisfaction probability, and contextual information.
[0159] The recognition module (140) recognizes the external environment by considering user-specified parameters, significance level, maximum conditional dissatisfaction probability, and contextual information, and transmits the recognition results to the control unit (180). The auxiliary data test results corresponding to the artificial intelligence model parameters are transmitted to the control unit (180).
[0160] The control unit (180) transmits user-specified parameters, significance levels, maximum condition dissatisfaction probability, and context information to the parameter and test result storage unit (165), receives recognition results from the recognition module (140), solves an optimization problem according to the recognition results under user-specified parameters, calculates an optimal control command, and uses this to control the target device (190).
[0161]
[0162] The system design and initialization process required when applying a control device utilizing the recognition results of a recognition module according to a preferred third embodiment of the present invention are described.
[0163] First, all inaccurate variables are eliminated from the system dynamics model. This is done using the initial values and control commands for these variables. All inaccurate initial values and model parameters throughout the model are considered as targets for recognition by the recognition module. Since the initial values and model parameters remain constant, once the elimination process is complete, only the variables and control commands whose values are known accurately remain in the dynamics model, excluding the parameters and initial values that are targets for recognition by the recognition module at the start of control.
[0164] The constraints that the control command of the control device utilizing the recognition result of the recognition module according to the third preferred embodiment of the present invention must satisfy are classified as follows.
[0165] 1) Constraints that are determined and known regardless of the target environment
[0166] 2) Constraints that depend on the target environment but are of low importance and therefore do not require probabilistic satisfaction guarantees.
[0167] 3) Constraints that depend on the target environment and require probabilistic satisfaction guarantees
[0168] In a third preferred embodiment of the present invention, only the case where the target environment for recognition and the recognition result of the recognition module (140) for the environment, on which the constraint of the above 3) depends, are both discrete and can have only a finite number of values is dealt with.
[0169] An optimization problem for determining optimal control commands based on the recognition results using a dynamic model is formulated as shown in Equation 10 below. Meanwhile, the constraints corresponding to 3) above are left undefined. That is, unlike Equation 5 in the second preferred embodiment of the present invention, some constraints in Equation 10 are undefined.
[0170]
[0171] Here, x and z are continuous and discrete control variables, respectively, o is the recognition result of the recognition module, and r is a user-specified parameter. This optimization problem has an objective function constraints By minimizing under, And it does not correspond to the constraints of the above 3) is determined by the system designer, taking into account the dynamic model and required control specifications.
[0172] Finally, a loss function L(x,z,θ,r) is defined as an evaluation of a control command of a control device according to a preferred third embodiment of the present invention.
[0173] Since the loss function is a post-evaluation of control commands rather than a judgment made during the control process, the actual environment θ is used instead of the recognition result o of the recognition module. This control system ultimately operates with the goal of minimizing the loss function.
[0174]
[0175] Figure 10 is a flowchart of the learning process of the recognition module when safety is ensured using auxiliary data according to a preferred third embodiment of the present invention.
[0176] Step 1000 is the first step of the model learning process, in which the operator of the model learning process transmits user-specified parameters, learning conservative parameters, and maximum conditional dissatisfaction probability (the learning conservative parameters and maximum conditional dissatisfaction probability can be set collectively for all the above 3) types of constraints, but can also be set for each constraint.) to the recognition module (140), and transmits user-specified parameters, significance level, maximum conditional dissatisfaction probability, and approximation parameters λ and β to the approximation loss function calculation unit (155). (λ is a scalar and β is the number of constraints in the optimization problem.) is a vector with the dimension of .)
[0177] At step 1010, the context information provision unit (160) transmits context information to the recognition module (140) and the approximate loss function calculation unit (155).
[0178] At step 1020, test data is input from the second data supply unit (120) to the test unit (130), and auxiliary data is input from the first data supply unit (110) to the test unit (130).
[0179] At step 1030, the recognition module (140) performs recognition on both the test data and the auxiliary data, and the test unit (130) transmits the test data (meaning accurate information (θ) about the environment) to the approximate loss function calculation unit (155).
[0180] When the test of the recognition module (140) by the test unit (130) is completed at step 1040, the recognition module (140) transmits the recognition result to the approximate loss function calculation unit (155).
[0181] At step 1050, the approximate loss function calculation unit (155) calculates the approximate loss function as follows.
[0182] First, the constraints of type 3) that were left undecided are determined as follows in a similar manner to the first embodiment of the present invention.
[0183] (1) o cr o is the part required for the constraint condition of type 3) above. cr These are the possible values for each j. If we assume that this is the case, There are only a finite number of these values. Let's put it this way. Also, let's put it as a discrete variable o cr A continuous variable w that is calculated as an intermediate step to calculate cr , that o cr A function that calculates o cr =f cr (w cr ) is placed.
[0184] (2) Each θ is as follows cr Find the maximum and minimum values of the probability of each recognition result under the given learning conservative parameter ξ.
[0185]
[0186] θ cr is the part related to the constraint of type 3) above among θ, and is θ cr This is one of the possible values N t is the total number of test data, N for condition g g means the number of auxiliary data satisfying g, and the superscript t k means the kth test case of auxiliary data.
[0187] (3) Each θ under each recognition result according to the learning conservative parameter given below cr Find the maximum probability of this occurrence.
[0188]
[0189] N here e is possible θ cr is the number of, Is It refers to the probability of contextual information.
[0190] (4) Determine the constraints of type 3) above that were left undecided as follows.
[0191]
[0192] Here, M is a sufficiently large number, and r t,i is the maximum probability of conditional dissatisfaction.
[0193] Now that all the constraints that were previously unknown have been determined, the optimization problem is complete, and the approximate loss function can be defined as follows using the same method as the second embodiment of the present invention.
[0194]
[0195] When determining constraints in this way, if the auxiliary data (a set of test cases) is sufficiently good, the true value of the performance of the recognition module is known, and accordingly, the constraints are guaranteed to be more conservative than those obtained based on Bayes' theorem, effectively solving the problem of overfitting to the auxiliary data.
[0196] Now for learning The (approximate) slope can be obtained as follows.
[0197] 1) Approximate gradient for elements that are not related to the constraints of type 3 among the discrete elements of recognition result o: As before, f(w i ) is a suitable softened probability distribution p(o di :w i ) can be replaced with. The expected approximation loss function and its approximate gradient are obtained as follows.
[0198]
[0199] O here di is o di It means a set of possible values, and in this invention, O di We only deal with the case where all of them are finite sets.
[0200] 2) Approximate gradient for a continuous element of recognition result o: As before, one continuous element o ci The approximate gradient for can be defined as follows for sufficiently small ρ:
[0201]
[0202] 3) The (approximate) gradient of the special case where the following values are defined for the continuous elements of the recognition result o:
[0203]
[0204] This value can be used as an (approximate) slope, and may be an exact slope depending on the situation. Here
[0205]
[0206] am.
[0207] 4) Approximate gradient for elements related to constraints of type 3 (necessary to check whether constraints are satisfied) among the discrete elements of recognition result o:
[0208]
[0209] Here o ncr Silver recognition result o middle o cr It means the part excluding, It means, silver cast When smoothing (same for -ξ), a specific learning conservative parameter is given. This means the probability of this coming out.
[0210] If this calculation is too complex or time-complex, It can also be calculated by approximating it with a continuous function.
[0211] Approximate loss function at step 1050 The gradient value of the received approximate loss function is returned to the recognition module (backpropagation begins). It is preferable that the recognition module (140) use the gradient value of the received approximate loss function to set or improve the artificial intelligence model parameters.
[0212] When the conditions initially specified by the operator are satisfied, learning is stopped, and the recognition module (140) calculates the parameters of the artificial intelligence model, and the approximate loss function calculation unit (155) calculates the aggregated test results (actual environment information for each test case). o The recognition result of the final model when the learning conservative parameter is 0 cr) is transmitted to the parameter and test result storage unit (165) and stored. If the condition is not satisfied, the process returns to step 1020 and repeats this process.
[0213]
[0214] Figure 11 is a flowchart of a control method for ensuring safety using auxiliary data according to a preferred third embodiment of the present invention.
[0215] In Fig. 11, it is seen that a test is performed on auxiliary data that was not used in pre-learning, and the results are stored in the parameter and test result storage unit (165).
[0216] The process of storing in the parameter and test result storage unit (165) is the same as that of FIG. 4 according to the first preferred embodiment of the present invention. At this time, the control unit (180) of FIG. 4 can be viewed as the parameter and test result storage unit (165) of FIG. 11.
[0217] Figure 11 was created based on robust control based on significance level, but conservative control based on learning conservative parameter can also be used instead of robust control based on significance level in general control.
[0218] Robust control based on significance level is more convenient than control techniques based on learning conservatism parameters, but it cannot be used for learning because it causes overfitting.
[0219] At step 1100, the control unit (180) receives context information from the context information provider (160). In addition, it receives user-specified parameters, significance level, and maximum conditional dissatisfaction probability from the user.
[0220] At step 1110, the control unit (180) transmits context information, user-specified parameters, significance level, and maximum conditional dissatisfaction probability to the parameter and test result storage unit (165).
[0221] At step 1120, the parameter and test result storage unit (165) selects suitable artificial intelligence model parameters based on the contextual information, user-specified parameters, significance level, and maximum conditional dissatisfaction probability, and transmits the artificial intelligence model parameters, together with the contextual information, user-specified parameters, significance level, and maximum conditional dissatisfaction probability, to the recognition module (140). In addition, the aggregated auxiliary test data results are transmitted to the control unit (180).
[0222] At step 1130, the recognition module (140) recognizes the external environment by considering contextual information, user-specified parameters, significance level, and maximum conditional dissatisfaction probability, and transmits the recognition result to the control unit (180).
[0223] At step 1140, the control unit (180) estimates the maximum probability of each situation from the context information, the aggregated auxiliary data test results, the recognition results, and the significance level, as in Equations 17 and 18.
[0224]
[0225]
[0226] Here o cr is a necessary part of the constraint condition of type 3) above, and θ cr is the part related to θ, and are θ respectively cr and o cr This is one of the possible values. means the number of auxiliary data satisfying k. is possible θ cr is the number of, Is It refers to the probability of contextual information.
[0227] Based on the maximum probability and maximum condition dissatisfaction probability of each situation estimated at step 1150, the risk limit constraints of the optimization problem (in the third preferred embodiment of the present invention, the constraints of the above 3) are determined as follows.
[0228]
[0229] Here, M is a sufficiently large number.
[0230] At step 1160, based on the recognition result, the risk limit constraints are determined according to equation 19 by reflecting the user-specified parameters, solving the optimization problem equation 10, and the control command is determined according to the optimal solution, and transmitted to the controlled device (190).
[0231]
[0232] The following examples illustrate the scope of the present invention. As an example of its application, the present invention can be used in autonomous vehicles or autonomous robots to recognize external signs, sounds, signals, or user instructions and then control them accordingly. Furthermore, the present invention can be utilized in automated financial product trading, such as stocks and bonds, by receiving and recognizing financial information, considering user preferences, and ensuring that principal is not lost beyond a certain probability. Beyond the two examples above, the present invention can be applied to a variety of other areas.
[0233] The term '~ unit' used in this embodiment means a software or hardware component such as an FPGA (field-programmable gate array) or an ASIC, and the '~ unit' performs certain roles. However, the '~ unit' is not limited to software or hardware. The '~ unit' may be configured to be on an addressable storage medium and may be configured to play one or more processors. Thus, as an example, the '~ unit' includes components such as software components, object-oriented software components, class components, and task components, processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functionality provided within the components and '~ units' may be combined into a smaller number of components and '~ units' or further separated into additional components and '~ units'. Additionally, components and '~parts' may be implemented to regenerate one or more CPUs within a device or secure multimedia card.
[0234] All of the functions described above can be performed by a processor such as a microprocessor, controller, microcontroller, ASIC (Application Specific Integrated Circuit), etc., according to software or program code coded to perform the functions described above. The design, development, and implementation of the code will be apparent to those skilled in the art based on the description of the present invention.
[0235] Although the present invention has been described with reference to specific embodiments, those skilled in the art will appreciate that various modifications and variations can be made to the present invention without departing from the technical spirit and scope of the present invention. Therefore, the present invention is not limited to the above-described embodiments, and includes all embodiments within the scope of the following claims.
Claims
1. A step in which the control unit estimates the probability that the actual situation corresponds to the i-th situation based on the recognition result in which the recognition module recognizes the external environment using the aggregation result generated by the aggregation module and the context information received from the context information provision unit; and The control unit includes a step of determining an optimal control command such that the sum of probabilities of situations in which the control command does not satisfy the constraint condition is less than the maximum condition dissatisfaction probability, A control method utilizing the recognition result of a recognition module, wherein the above contextual information is a contextual probability that the actual reality is the i-th situation.
2. In paragraph 1, A step in which the above recognition module recognizes auxiliary data and generates a recognition result; and A control method utilizing the recognition result of a recognition module, characterized in that the aggregation module further includes a step of aggregating the recognized auxiliary data and the recognition result to generate an aggregate result.
3. In paragraph 2, A control method utilizing the recognition results of a recognition module, characterized in that the probability of corresponding to the i-th situation is estimated by further considering the significance level, which is the maximum allowable limit of the probability that the estimation range of the recognition module performance based on the aggregated results does not include the actual value due to the limited number of the auxiliary data.
4. A computer-readable recording medium having recorded thereon a program for executing the method of any one of clauses 1 to 3 on a computer.
Citation Information
Patent Citations
Generation device, program, recognition system, and generation method
JP2018147391A
Accident occurrence probability estimation device, and control method and program of accident occurrence probability estimation device
JP2023023481A
Parking control system, parking control method, and parking control program
JP2023061755A
Driving path updating apparatus based on driving information and environment recognition and method thereof
KR1020170098070A
Laundry Treating Apparatus Control System by Voice Recognition and operating Method the same
KR102498007B1
