JP7370924B2 - Optimization device and optimization method - Google Patents

Description

The present invention relates to a parameter optimization technique.

In recent years, as the performance of computers has improved, the scope of parameter optimization using computers has been expanding more and more. However, when the evaluation index (hereinafter referred to as evaluation function) depends on a plurality of parameters in a complicated manner, it is often a problem that optimization of the parameters still requires a huge amount of time. In particular, in situations where the specific function form of the evaluation function is not known, the value of the evaluation function when the parameter values are determined (hereinafter referred to as the evaluation value) can be found, but the differential value cannot be found. If it is necessary to calculate multiple evaluation values under each value and directly compare them, and the time required to calculate each evaluation value cannot be ignored, the length of calculation time becomes significant.

Therefore, instead of randomly generating parameter values to be evaluated (hereinafter referred to as search points) each time, by generating the next search point by considering the evaluation value of the previously calculated search point, we can efficiently calculate the optimal parameter value. Optimization methods that search for are widely used. For example, in a group of optimization methods known as distribution prediction algorithms, search points are generated based on a probability distribution explicitly constructed from evaluation value information (hereinafter referred to as search point generation distribution), which improves search efficiency. We are trying to make this happen.

However, one problem with such optimization methods is that search performance deteriorates when the evaluation function has a flat region that hardly changes with changes in a certain parameter.

In Non-Patent Document 1, CMA-ES (Covariance Matrix Adaptive Evolution Strategy: In a distribution prediction algorithm called Covariance Matrix Adaptation Evolution Strategy, in order to reduce the deterioration of search performance for poorly conditioned evaluation functions, we limit the search target parameters to a small number of randomly selected parameters each time. is suggesting. Here, the ill-conditioned function refers to a multivariable function whose degree of curvature differs greatly depending on the direction, and functions with flat regions are included in the ill-conditioned function. In addition, in Patent Document 1, by introducing a decision device configured on a learning basis based on past search points and their evaluation values, and determining whether to perform evaluation value calculation for each search point, It is proposed to eliminate unnecessary evaluation value calculations and reduce calculation time.

Japanese Patent Application Publication No. 2019-192160

Koki Shimizu, Junpei Komiyama, Masashi Toyota, “Stochastic dimension selection CMA-ES for high-dimensional ill-conditioned optimization problems”, DEIM Forum 2019 A4-3

In the method proposed in Non-Patent Document 1, since the selection of search target parameters is random, search parameters cannot be selected according to the characteristics of individual adverse conditions. In particular, it has been pointed out in Non-Patent Document 1 that search performance deteriorates compared to conventional CMA-ES for evaluation functions that are not under bad conditions.

Patent Document 1 assumes that a determiner is constructed on a learning basis. Therefore, in principle, it is possible to deal with various causes (including flatness) that cause deterioration of search performance. However, challenges include the large amount of data that must be retained for learning and the large learning costs.

The present invention was made in view of the above problems, and aims to provide a method for realizing efficient search by avoiding deterioration in search performance caused by flatness among other unfavorable conditions. do.

A preferred aspect of the present invention is an optimization device that optimizes parameter values. This device includes an input section that receives a parameter to be searched and an evaluation function serving as its evaluation index, an optimization calculation section that calculates an optimal value of the parameter based on the evaluation function, and an output section that outputs the optimal value. and, including. The optimization calculation unit includes a search point generation unit that generates a search point that is a value of a parameter to be evaluated from a search point generation distribution, and an evaluation value calculation unit that calculates an evaluation value of the search point based on the evaluation function. , a distribution shape updating unit that updates the search point generation distribution based on the evaluation value; a distribution shape holding unit that retains a quantity characterizing the search point generation distribution as time series information; The search parameter selection unit selects parameters to be searched for, and the termination determination unit determines termination based on predetermined termination conditions.

Another preferred aspect of the present invention is an optimization method that is executed by an information processing device including an input device, an output device, a processor, and a storage device, and optimizes parameter values. This method includes a first step of receiving a parameter to be searched and an evaluation function serving as its evaluation index, a second step of generating a search point, which is the value of the parameter to be searched, from a search point generation distribution, and a second step of receiving the parameter to be searched and an evaluation function serving as its evaluation index. a third step of calculating an evaluation value of the search point based on the evaluation value; a fourth step of updating the search point generation distribution based on the evaluation value; and a fourth step of updating the search point generation distribution based on the evaluation value. a sixth step of selecting parameters to search based on the time series information;
Execute.

According to the present invention, efficient search is possible even when the evaluation function includes a flat area that causes deterioration in search performance.
Problems, configurations, and effects other than those described above will be made clear by the following description of the embodiments.

It is a block diagram showing an example of functional composition of an optimization device of an example. FIG. 2 is a block diagram showing an example of a hardware configuration of an optimization device according to an embodiment. It is a flowchart which shows an example of processing of the optimization device of an example. It is an explanatory diagram showing an example of processing of a distribution prediction type algorithm. FIG. 3 is an explanatory diagram showing an example of redundant calculations due to flatness. It is an explanatory diagram showing an example of processing of search parameter selection of an optimization device of an example. 7 is a flowchart illustrating an example of detailed processing of a search parameter selection unit. FIG. 6 is an explanatory diagram showing an example of processing when there is a correlation between parameters. FIG. 7 is a diagram illustrating an example where the flatness of the evaluation function is local. It is an explanatory view showing an example of a characteristic response of an optimization device of an example. 7 is a perspective view showing an example of an output device that displays the degree of search progress in Example 2. FIG.

Embodiments of the present invention will be described below with reference to the drawings. Note that this embodiment is merely an example for implementing the present invention, and does not limit the technical scope of the present invention.

In the configuration of the invention described below, the same parts or parts having similar functions may be designated by the same reference numerals in different drawings, and overlapping explanations may be omitted.

When there are multiple elements having the same or similar functions, the same reference numerals may be given different subscripts for explanation. However, if there is no need to distinguish between multiple elements, the subscript may be omitted in the explanation.

In this specification, etc., expressions such as "first," "second," and "third" are used to identify constituent elements, and do not necessarily limit the number, order, or content thereof. isn't it. Further, numbers for identifying components are used for each context, and a number used in one context does not necessarily indicate the same configuration in another context. Furthermore, this does not preclude a component identified by a certain number from serving the function of a component identified by another number.

The position, size, shape, range, etc. of each component shown in the drawings etc. may not represent the actual position, size, shape, range, etc. in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings or the like.

The publications, patents, and patent applications cited herein are incorporated in their entirety.

Elements expressed in the singular herein shall include the plural unless the context clearly dictates otherwise.

The following is employed as one of the typical configurations of the embodiments described in detail below. This example is an optimization device that inputs a parameter to be searched, its evaluation function, initial values of the parameter, and hyperparameter values, and outputs a parameter value that maximizes the evaluation value. This device includes a search point generation section that generates search points from a certain search point generation distribution, an evaluation value calculation section that calculates the evaluation value of the search point, and a distribution that updates the search point generation distribution based on the evaluation value. A shape updating unit, a distribution shape holding unit that retains a quantity that characterizes the distribution (for example, a variance value), and a search parameter that selects whether to perform a search for each parameter based on the information on the distribution shape. It includes a selection section and an end determination section that determines whether or not the search is complete. Furthermore, when the number of search parameters changes as a result of the selection process, the search parameter selection section also performs processing to update hyperparameters of the optimization method that depend on the number of search parameters to appropriate values.

According to the above configuration, it is possible to estimate the shape of the evaluation function by referring to the time series information of the shape of the search point generation distribution stored in the distribution shape holding section, and the estimation result is obtained in the search parameter selection section. It is possible to select search parameters according to the

This embodiment can be used, for example, when optimizing the threshold for feature point extraction in object recognition. For example, in order for a picking robot used at a logistics site to be able to recognize the positions and postures of various items, it is effective to perform optimization processing on parameters such as threshold values. By adopting this embodiment, it is possible to improve the accuracy and speed of recognition while increasing the efficiency of parameter optimization processing.

FIG. 1 is a block diagram showing an example of the functional configuration of this embodiment. As shown in FIG. 1, the optimization device 1 of this embodiment registers an evaluation function serving as an evaluation index of an optimization target, parameters on which the evaluation function depends, and initial values of the parameters and values of hyperparameters. It consists of an input section 11, an optimization calculation section 12 that searches for a parameter value that maximizes the value of the evaluation function by a loop process described later, and an output section 13 that outputs the resulting parameter value.

Furthermore, the optimization calculation unit 12 includes a distribution shape holding unit 121 that holds quantities characterizing the shape of the search point generation distribution, selects search parameters based on this information, and updates hyperparameters that depend on the number of search parameters. A search parameter selection unit 122, a search point generation unit 123 that generates a search point from a search point generation distribution for the parameter selected therein, and an evaluation value calculation unit 124 that calculates an evaluation value at the search point. An end determination unit 125 determines whether to end the search based on the target value of the evaluation value and the maximum number of evaluation value calculations, and if the search does not end based on the determination, the end determination unit 125 performs the search based on the search point and its evaluation value. The distribution shape updating section 126 updates the point generation distribution and stores the quantity characterizing the shape in the distribution shape holding section 121.

Here (and below) we will explain an example of an optimization method in which search points are generated from an explicit probability distribution, but if the search point generation method can be reduced to sampling from a probability distribution, then this can be included in the embodiment. In the configuration of the optimization calculation unit 12 described above, the loop process continues until the termination condition of the termination determination unit 125 is satisfied, and for each loop process, the amount characterizing the search point generation distribution is accumulated in the distribution shape storage unit 121 in time series. It will be done. Therefore, the search parameter selection unit 122 can utilize time-series information about the shape of the search point generation distribution as a criterion when selecting search parameters. However, as will be described later, data that is not used by the search parameter selection unit 122 may be discarded. It is also relatively memory efficient, since instead of storing all search points and evaluation values in each loop, it only needs to store a relatively small number of quantities that characterize the distribution shape extracted from them. is mentioned as one of the features of this embodiment.

The method for updating the search point generation distribution in the distribution shape updating unit 126 is not required to be unbiased. Here, unbiasedness in distribution update means that when there is no correlation between the search points and their evaluation values (for example, uniform, random, etc.), the distribution update is such that the fluctuation of the distribution due to the update becomes zero on average. refers to the properties of If there is no unbiasedness, the value of the optimal parameter to be found will be affected by the initial value setting of the search parameter, so it is generally undesirable, although it depends on the purpose.

Here, by retaining the information accumulated in the distribution shape retaining unit 121 as a result of executing this embodiment even after the optimization calculation is completed, the following effects can be obtained. In other words, when you want to perform an optimization calculation similar to an optimization calculation that has already been performed, when setting the parameters you want to search for, their initial values, and hyperparameter values in the input unit 11, you need to store them in the distribution shape storage unit 121 described above. You can utilize the information provided. For example, as will be described later, it is possible to infer the existence and range of a flat region of the evaluation function from the information in the distribution shape holding unit 121, so it is possible to remove in advance the corresponding parameters of low importance to be searched. It is possible to limit its domain.
Up to this point, in this embodiment, it has been assumed that the initial values of parameters and the values of hyperparameters to be registered in the input unit 11 are directly specified from the outside, but they may also be determined by internal processing. For example, it is conceivable to randomly select initial values of parameters from the domain when there is no prior knowledge.

FIG. 2 shows an example of a hardware configuration for realizing the functional configuration example of FIG. 1. As shown in FIG. The optimization device 1 of this embodiment has, for example, a processor 101, a memory 102, an auxiliary storage device 103, an input device 104, an output device 105, and a communication IF (Interface) 106, which are connected to an internal communication line such as a bus. It is composed of computers connected by 107.

The processor 101 executes a program stored in the memory 102 to realize functions other than the distribution shape holding section 121 of the optimization calculation section 12. The distribution shape holding unit 121 is mainly realized by the memory 102. However, if the amount of data to be held is large or if the obtained data is to be used for other similar optimization calculations, the auxiliary storage device 103 will also play this role. The memory 102 is, for example, a non-volatile storage element (for example, ROM (Read Only Memory)) for storing programs that do not need to be changed, and temporarily stores programs to be executed and data used during program execution. It consists of a volatile storage element (for example, RAM (Random Access Memory)) for On the other hand, the auxiliary storage device 103 includes a nonvolatile, large-capacity storage device such as a magnetic storage device (HDD (Hard Disk Drive)), and stores programs executed by the processor 101 and data used during program execution. do. Due to the above factors, the optimization program is first read out from the auxiliary storage device 103, loaded into the memory 102, and executed by the processor 101, for example.

The input device 104 is a device such as a keyboard or a mouse that receives input from an operator, and enables input operations to the input unit 11 of the optimization device. The output device 105 is a device, such as a display or a printer, that outputs the results of program execution (for example, the output of the output unit 13) in a format that can be recognized by the operator. The communication IF 106 is a network interface device that controls communication between this optimization device and other devices.

As described above, in this embodiment, functions such as calculation and control are performed by executing programs stored in the memory 102 and the auxiliary storage device 103 by the processor 101, thereby performing predetermined processing in cooperation with other hardware. It is realized by working. A program executed by a computer or the like, its function, or a means for realizing that function may be called a "function," "means," "section," "unit," "module," etc. The search parameter selection unit 122, search point generation unit 123, evaluation value calculation unit 124, termination determination unit 125, and distribution shape update unit 126 shown in FIG. It is assumed that the data is stored in the auxiliary storage device 103. Note that functions equivalent to those configured using programs can also be achieved using hardware such as FPGA (Field Programmable Gate Array) and ASIC (Application Specific Integrated Circuit).

The above configuration may be configured by a single computer, or any part of the processor 101, memory 102, auxiliary storage device 103, input device 104, output device 105, and communication IF 106 may be connected via a network. It may also be composed of other computers.

FIG. 3 is a flowchart showing an example of processing performed by the optimization calculation unit 12 of this embodiment. In the optimization calculation of this embodiment, CMA-ES is partially applied. CMA-ES is a type of evolutionary computation, and is a multi-point search method that considers the solution to the target problem as a pseudo-biological individual and searches for a solution using that population (individual population). . As is well known, in CMA-ES, the next generation population is generated by mutation based on a normal distribution, and the covariance matrix of the normal distribution is updated using a mechanism called Covariance matrix adaptation.

Among the processes shown in FIG. 3, among the search point generation step S30, evaluation value calculation step S40, end determination step S50, and step S60, the conventional CMA-ES technology is applied to update the search point generation distribution. You may do so. Although CMA-ES is used in this example, various distribution prediction algorithms can be used instead of CMA-ES, including concepts such as real-valued genetic algorithms and ES (Evolution Strategy, ES) algorithms. You may.

The flow of the process will be explained according to the flowchart of FIG. As described above, the optimization calculation unit 12 performs loop processing, so the processing from the time when the k-th loop is entered will be explained. In the following, for the sake of explanation, the parameter to be searched for is represented by a d-dimensional vector x=(x ₁ ,..., x _d ). d is a natural number representing the number of parameters, and each parameter takes a continuous value or a discrete value. Further, the evaluation function is represented by F(x), and the search point generation distribution is represented by P(S ^(k) ). Here, S ^(k) is a set of quantities characterizing the distribution shape, and the number of elements thereof is arbitrary.

First, in step S10, the search parameter selection unit 122 selects search parameters. Information stored in the distribution shape holding unit 121 is used as the selection criterion. The information stored in the distribution shape holding unit 121 is the history of S ^(k) including the past. Here, if the number of loops k is small and the data is not stored in the distribution shape holding unit 121 or is insufficient, the process moves to step S40. If sufficient data exists in the distribution shape holding unit 121, parameters to be searched are selected. Let I ^(k) be the index set of parameters to be searched, and J ^(k) be the index set of parameters not to be searched.

In the next step S20, the search parameter selection unit 122 first determines whether the number of search parameters has changed since the previous loop processing (k-1th loop). If |I ^(k) |=|I ^(k-1) |, the process moves to step S30 without doing anything. In the case of |I ^(k) |≠|I ^(k-1) |, the process moves to step S21, and the search parameter selection unit 122 updates the hyperparameter.

For example, in the case of CMA-ES, there are recommended values for the number of search points generated in step S30 (described later), the learning rate at the time of updating the distribution in step S60, etc. that make the search more efficient depending on the number of search parameters. After changing such hyperparameters to values corresponding to the number of search parameters, the process moves to step S30.

In order to update the hyperparameters, a data table that can be referenced by the search parameter selection unit 122 is prepared in the memory 102. The contents of the data table may store, for example, recommended values for the number of search points and the learning rate at the time of updating the distribution, corresponding to the number and range of search parameters.

The hyperparameter can also be set to a fixed value without being updated, and in that case, step S20 and step S21 can be omitted. However, the hyperparameter update process here can be expected to further improve search performance.

In step S30, the search point generation unit 123 generates a search point. The search point can be expressed by the following equation (1).

Here, x _i;n ^(k) represents the i-th component of the n-th search point generated in the k-th loop. Moreover, x ^{^} _i;n ^(k) is a value for a parameter for which no search is performed. Its value is not important as long as flatness is considered. This is because being flat means that the value of the evaluation function does not change even if you change the parameter value in that direction, and x ^{^} _{i; n} ^(k) is not related to the value of the evaluation function. , this is because the result does not change for any value. For example, it is conceivable to adopt a component corresponding to the center position of P as the value of x ^{^} _i;n ^(k) . In this way, according to equation (1), for the dimension belonging to I ^(k) , search points whose values have been changed from the search point generation distribution P(S ^(k) ) are selected, and the dimension belonging to J ^(k) is For , set it to an arbitrary fixed value.

Subsequently, in step S40, the evaluation value calculation unit 124 calculates the evaluation value at each search point. In other words, F(x _;n ^(k) ) is obtained for all n. The time it takes to calculate this evaluation value is the main reason why the calculation time required to find the optimal parameters is enormous. Here, in step S10, the i component of the unselected parameter is treated as a constant, so the calculation time is shortened.

In the next step S50, the end determination unit 125 determines the end of the search. The specific end determination can be set depending on the purpose. For example, it may be determined whether the number of loops k exceeds a certain maximum number of repetitions or whether the maximum evaluation value max(F(x)) exceeds a certain target value. If the termination condition is satisfied, the loop process is exited and the determined optimal parameter value argmax(F(x)) is passed to the output unit 13. On the other hand, if the termination condition is not satisfied, the process moves to step S60.

In step S60, the distribution shape updating unit 126 updates the search point generation distribution P based on the search points generated in the k-th loop and their evaluation values. Accordingly, S ^(k) is updated, and after storing necessary elements in the updated value S ^(k+1) in the distribution shape holding unit 121, the process returns to step S10.

A notable feature of the processing of this embodiment as described above is that the time-series information of the distribution shape stored in the distribution shape storage unit 121 is used to limit the parameters to be searched. Another feature is that hyperparameters are updated in accordance with parameter limitations. In order to explain its behavior and effects more specifically, below, we will set the number of parameters to be searched to 2 (that is, d = 2), and set the search point generation distribution P to a multivariate characterized by the center position and covariance matrix. Assume normal distribution. Then, each related quantity will be expressed by the following equation (2).

Here, μ ^(k) is the center position of P in the k-th loop, and C ^(k) is the covariance matrix of P in the k-th loop.

Using FIG. 4, we will first explain typical search point generation, evaluation value calculation, and update of search point generation distribution in the distribution prediction algorithm, which is the main target of this embodiment, when there is no search parameter selection process. Explain.

In (a1) of FIG. 4, a search point generation distribution which is a normal distribution and eight search points generated therefrom are described. For each search point, those with large evaluation values are marked as white points, and those with small evaluation values are marked as black points. The search point generation distribution modified based on the evaluation value is shown in (b1) together with the search points in (a1). As can be seen from (b1), the search point generation distribution is modified so that search points with large evaluation values are likely to be generated. (a2) shows the search point generation distribution of (b1) and eight new search points generated therefrom. As before, the result of updating the distribution based on the evaluation value is (b2). As can be seen from this flow, it can be seen that the search point generation distribution is gradually transformed into a distribution that generates search points with as large an evaluation value as possible. Here, it should be noted that FIG. 4 is a simple image diagram, and a more sophisticated updating method is adopted in the actual algorithm.

Next, using FIG. 5, the influence when a flat region exists in the evaluation function in the above-mentioned optimization method will be explained. In FIG. 5, we are considering an evaluation function that is convex upward in the _x1 direction but flat in the _x2 direction, as shown in (a). Consider a situation in which evaluation values are calculated for five search points generated from the search point generation distribution, as shown in (b). Here, since the value of the evaluation function does not change in the _x2 direction, the evaluation function can be substantially regarded as a one-variable function. In fact, as shown in (c), even if the evaluation value is calculated by projecting the value of x ₂ at all search points onto the value μ ₂ ^(k) of the corresponding component at the center position of the normal distribution, the result will not change. Then, the number of search points generated in each loop has a recommended value according to the number of search parameters, and if the number of recommended search points is 5 and 3 for 2 and 1 search parameters, respectively, (d) By limiting the number of search points to three, more efficient search becomes possible. Here, it should be noted that the width of the distribution in the _x2 direction does not change in subsequent calculations on average due to unbiasedness.

So far, we have seen that if there is a parameter that does not change the evaluation value and there is a flat region in the evaluation function, you can perform an efficient search by taking this into consideration. It must be detected during the search. Here, it is possible to investigate the flatness of the evaluation function in advance, but as will be described later, this is generally not practical. In this embodiment, flatness can be detected during the search.

Using FIG. 6, a description will be given of how flatness is detected during the search. Assume that the situation is as shown in (a) in FIG. 6 at the k−r+1th loop. Suppose that after r times of loop processing, the result is as shown in (b). In this example, it is assumed that the standard deviations of the normal distribution in the x ₁ and x ₂ directions up to before r loops are stored in the distribution shape storage unit 121. Here, due to unbiasedness, if the fluctuation of the standard deviation is small or less than a certain small threshold, it can be determined that the evaluation function is flat or random in the corresponding direction. For example, the determination condition can be expressed by the following equation (3).

Here, ε is a threshold value, and ΔS _i ^(k;r) is a quantity representing the degree of variation of the i-direction standard deviation from the k−r+1th loop to the kth loop. The latter can be defined, for example, as in equation (4) below.

Here, S _i ^(k) represents the i-direction standard deviation at the k-th loop. According to the above criteria, parameters that are excluded from the search target in a certain loop will not be searched after that, but considering the risk of falling into a local solution, it is necessary to decide whether to search for such parameters again. Additional judgments may be added. For example, the determination can be defined as in equation (5) below.

Here, c is a constant greater than 1. This formula represents a process in which all parameters excluded from the search are returned to the search target when fluctuations in the corresponding standard deviations of all parameters to be searched become small to a certain extent. Although the addition of such a determination can reduce the risk of falling into a local solution, it should be noted that the time required for the search increases.

It should also be noted that the above-mentioned indicators and criteria for the amount of variation are not limited to those listed above. In addition, when search parameters are selected based on the criteria described above, significant fluctuations in the distribution generally occur during searches in areas that are not flat, so it is possible to maintain the search performance without selection processing. I would like to emphasize this as an important feature. In any case, the method described above can determine parameters for which a stable improvement in the evaluation value cannot be expected compared to other directions even if further search is performed. In this embodiment, such processing can be realized by using the distribution shape holding section 121 and the search parameter selection section 122.

The right side of (c) in Figure 6 shows the case where the above processing is performed and the search is not performed in the _x2 direction after the kth loop, and the left side is the case where the above processing is not performed. This shows how the search is being carried out. Finally, in this example, the standard deviation held in the distribution shape holding unit 121 is for r loops, and it is possible to keep the amount of data to be held relatively small.

FIG. 7 is a diagram showing a detailed flow of step S10 executed by the search parameter selection unit 122 in order to realize the principle shown in FIG. 6. Solid line arrows indicate the flow of processing, and dotted line arrows indicate the flow of data. Here, i is the index of the parameter, and the initial value is i=1. In step S10, based on time-series information on the distribution shape of the search point generation distribution stored in the distribution shape holding unit 121, for example, the variance value of the distribution, parameters whose distribution shape changes are greater than or equal to a threshold are selected.

In step S11, the search parameter selection unit 122 reads the history of the standard deviation in each direction of the normal distribution up to a predetermined r loops before, from the distribution shape holding unit 121, that is, the history of the information S ^(k) .

In step S12, the search parameter selection unit 122 measures the change in the distribution shape by calculating the change in the standard deviation of the distribution in the i direction using, for example, equation (4).

In step S13, the search parameter selection unit 122 uses, for example, equation (3) to extract a parameter whose standard deviation is smaller than a threshold, estimates that the parameter has flatness, and excludes it from the search target. Here, excluding a parameter from search targets means performing subsequent processing as a constant. In this case, any constant can be selected as long as the parameters are completely flat. Furthermore, when flatness is estimated only within a predetermined range, a constant may be selected within the range where flatness is estimated. For example, since the flat range is estimated to be about the width of the search point generation distribution (about the distance of the standard deviation from the center of the distribution), values within that range may be selected. Since the center of the distribution is always within this range, in practice it is sufficient to select the center value. That is, parameters not selected by the search parameter selection unit 122 are fixed to values corresponding to the center value of the search point generation distribution.

In step S14, the search parameter selection unit 122 determines whether processing has been completed for all parameters. For example, the end determination is made by determining whether i has reached the maximum index number of the parameter. If not completed, the next parameter is processed in step S15.

When the processing has been completed for all parameters, the search parameter selection unit 122 uses, for example, equation (5) to determine whether the variation in standard deviation is the threshold value for all parameters or a predetermined proportion of parameters to be searched in step S16. Determine whether it is smaller than. As a result of the determination, if the variation in standard deviation is smaller than the threshold value, it is determined in step S17 whether or not to search again for the direction excluded from the search target. At this time, all excluded parameters may be restored as described above. Alternatively, the excluded parameters may be shown to the operator so that the operator can select the parameters to be restored.

Note that, as already stated, step S16 and step 17 can be omitted.

In the explanation of Figure 6, the explanation is based on the assumption that the two parameters are independent, but if they are correlated with each other, it is more appropriate to diagonalize the covariance matrix and focus on its eigenvalues. flat regions of the evaluation function can be detected.

This will be explained using FIG. In FIG. 8, the evaluation function shown in FIG. 5(b) is considered to be tilted 45 degrees to the left, and the straight line x ₁ =x ₂ is in a flat direction. Assume that the search progresses from the state shown in FIG. 8(a) and reaches a state as shown in FIG. 8(b). As mentioned above, considering the standard deviation in the x ₁ and x ₂ directions (the square root of the diagonal component of C ^(k) ), both are changing, so they are not excluded from the search target. However, when looking at the coordinate system (y ₁ , y ₂ ) rotated by 45 degrees, there is no change in the standard deviation in the y ₁ direction corresponding to the flat direction. In other words, it is clear that in order to more precisely detect flat areas for correlated parameters, it is necessary to pay attention to the fluctuations in the eigenvalues of the covariance matrix.

In order to implement such a function, if there is a quantity expressed by a second-order matrix among the quantities that characterize the shape of the search point generation distribution, the search parameter selection unit 122 saves it in the distribution shape storage unit 121. In order to extract the quantity, it is sufficient to have a function to perform matrix diagonalization processing. Alternatively, similar processing may be performed when storing the quantity characterizing the shape of the search point generation distribution in the distribution shape holding unit 121.

In this example, we are considering a two-dimensional covariance matrix, but it is also possible to extend it to higher dimensions. However, in the above process, matrix diagonalization is required in each loop process, so it is necessary to compare the time required for diagonalization and the resulting degree of efficiency to decide whether to actually use it. There is a need.

It is conceivable to detect a flat region of the evaluation function in advance, instead of detecting it during the search as in this embodiment, but the flatness of the evaluation function also changes depending on the range of values taken by the parameters. Therefore, in order to properly understand the flatness, a comprehensive investigation over the entire definition range of each parameter is required, which requires a corresponding amount of calculation cost.

FIG. 9 is a diagram illustrating an example where a flat region of the evaluation function exists locally for one parameter. As shown in this figure, this evaluation function is determined to be flat for region 91, but not for region 92.

Finally, the characteristic response of the optimization apparatus in this embodiment will be described using FIG. 10. In this embodiment, we focus on time-series changes in the distribution shape in order to detect parameters that are less important to search for further, but since this judgment relies on the unbiasedness of the distribution update, The distinction between flatness and randomness is not made. However, with regard to randomness, the importance of continuing the search is low in the sense that no significant difference in evaluation values can be found even if the search is continued in that area, similar to the case of flatness.

In the input unit 11 of this embodiment, it is necessary to provide the parameters to be searched and an evaluation function depending on them. When designing the evaluation function, external input is often required. For example, in fluid simulation, it is necessary to provide values for parameters such as fluid viscosity. Furthermore, in image recognition, if one wants to find a parameter that maximizes the recognition rate of a certain object, it is necessary to provide an image of the object in addition to information about the object. It is often possible to bring flatness or randomness to the evaluation function through artificial changes to such external inputs.

For example, in the image recognition example above, the evaluation function can be made random by adding large random noise to the input image, and the evaluation function can be flattened by making the features such as color of the input image uniform. There are cases. When an evaluation function designed through such processing is input to the optimization device 1 of this embodiment, in both cases, the necessity of the search is detected and the search is promptly terminated. For example, in an optimization method such as the gradient method, the behavior is similar in the flat case, but there is a noticeable difference in that in the random case, the calculation is unstable and the search does not end. In other words, it can be said that such response characteristics are noteworthy as a feature of this embodiment. Another characteristic response is that even if you add an arbitrary number of parameters that are completely unrelated to the evaluation function to the parameters you want to search, it will quickly be detected that they are unnecessary, so the search will take longer. It can be mentioned that it is almost constant regardless of the number of additions.

In this embodiment, the optimization device of the embodiment is applied to optimization of object recognition parameters of a picking robot. However, it should be noted that this embodiment describes an example of the effect of the optimization device of the embodiment on specific hardware, and does not limit its application. .

In recent years, due to the labor shortage, demand for autonomous work robots to replace manual labor at logistics and manufacturing sites has increased. In particular, picking robots that can grasp a target object and place it in a predetermined location are expected to be useful in a variety of situations. Here, in order for such a picking robot to correctly grasp an object, it must be able to accurately recognize the position, orientation, and type of the object. However, considering the operation in the field where a variety of objects are handled or the objects being handled change frequently, in order to maintain a high recognition rate, it is necessary to adjust the parameters related to object recognition processing every time the target object changes. , it is necessary to adjust it according to the object. In this embodiment, the optimization apparatus of the first embodiment is responsible for parameter optimization for each object. That is, the configuration of this embodiment is the same as that of the first embodiment, except that the picking robot has an object recognition and grasping function using the optimal parameters output to the output unit 13 of the optimization device. In the following, the input/output of the optimization device and its effects in this example will be specifically explained.

The object recognition function of the picking robot of this embodiment determines whether or not there is a target object in an image of a certain situation (hereinafter referred to as a scene image) taken with a camera, and if so, where and in what orientation it exists. The decision shall be made as to whether or not to do so. Generally, a program that implements this function has multiple parameters that can affect recognition performance. For example, in the case of pose estimation based on matching the feature points of the target object acquired in advance with the feature points of the scene image, parameters related to the acquisition of the feature points from the scene image (for example, the parameters that affect the recognition performance) (minimum permissible value of brightness gradient to be considered), feature point matching (for example, maximum permissible value of distance to be considered as a set), etc. The evaluation function in such parameter optimization is recognition performance, specifically the recognition rate of the target object with respect to the scene image. However, depending on the purpose, a term may be added such that the shorter the recognition processing time, the larger the value. In any case, since complex recognition processing is involved, the specific functional form of the evaluation function is unknown, and information on differential values cannot be used. Therefore, in the present case, optimal parameter search using a search point generation distribution (as targeted by this embodiment) is a powerful optimization method.

The input unit 11 of the optimization device receives information on the target object (for example, feature points and their feature amounts), a group of scene images, parameters related to object recognition processing and their initial values, and hyperparameter values. Here, it is generally difficult to pre-select whether or not each parameter is included in the search target, and it is desirable to register all parameters in the input unit 11 as initial parameters. However, the disadvantage of this is that unnecessary parameters may be included and search performance may deteriorate. The optimization calculation unit 12 searches for parameters that maximize recognition performance, but in order to calculate the evaluation value, it is necessary to perform recognition on a group of scene images each time, which generally takes a long time, so it is difficult to improve the efficiency of the search. The benefits are great. The optimal parameters outputted to the output unit 13 are sent to the control device of the picking robot via the output device or the communication IF, thereby realizing accurate gripping of the target article. Note that the optimization device 1 may be a part of the control device of the picking robot, or may be a separate and independent configuration.

The effects of this embodiment in the application example described above are as follows. As described above, in this embodiment, unnecessary input parameters may be included. Furthermore, since a plurality of parameters depend on each other in a complicated manner, there is a strong possibility that the evaluation function has a plurality of local flat regions. However, in the optimization device of this embodiment, parameters with low search importance are detected and removed during the search by the search parameter selection process, so that the influence of search deterioration can be reduced. Therefore, the teaching necessary for grasping the target object can be efficiently performed, and the picking robot can easily handle a wide variety of products.

In the present embodiment, in the configuration of the embodiment in Example 1, the output unit 13 has a function of receiving not only the optimal parameters obtained at the end of the optimization calculation but also information obtained during the search. It is characterized by being displayed externally in real time through an output device. As shown in the embodiment of Example 1, for each loop process, the distribution shape storage unit 121 stores the quantity characterizing the shape of the search point generation distribution during the search, and further stores the quantity characterizing the shape of the search point generation distribution during the search. Through the process 122, it is selected whether each parameter is a search target. The output unit 13 of this embodiment can receive such information in addition to the optimal parameters during the search. The hardware that implements the output device in this embodiment is not limited as long as it can constantly display received information in a form that can be visually recognized by the operator; for example, a display may be used.

FIG. 11 is a diagram showing an example of a hardware configuration for realizing this embodiment. In this example, a computer includes a processor 101, storage devices such as a memory 102, and an auxiliary storage device 103, a keyboard as an input device 104 for instructing execution of an optimization program, and a display 1101 as an output device 105. , is depicted. In this example, the information displayed on the display 1101 changes every moment during the optimization calculation. Specifically, parameters selected by the search parameter selection unit 122 and parameters not selected are displayed on the display 1101.

Additionally, a configuration has been added that allows reference to various information held by the distribution shape holding unit 121 in order to select input parameters in the input unit, set their domains and initial values, and set hyperparameters. It's okay.

By adding the above functions, for example, the following advantages are created compared to the embodiment of the first embodiment.

The first is that it becomes possible to estimate the degree of progress of the search more accurately than by estimating based on the number of initial input parameters. For example, even if the evaluation value calculation result in a certain loop process is the same when the number and type of parameters excluded from the search target are known and when they are not known, the remaining search will finish faster in the former case. You can make a judgment. Such a more accurate estimate can be useful for ending the search early, depending on the purpose. For example, if the maximum value of the evaluation function is unknown and it is not possible to use whether the maximum evaluation value has reached the target value to determine the end of the search, an additional termination condition may be used to determine whether all parameters are no longer included in the search target. can do. Furthermore, with this advantage, it is possible to change the optimization method (for example, the method of updating the search point generation distribution) depending on the number of search parameters.

Second, by displaying the quantities that characterize the shape of the search point generation distribution, similarity optimization calculations can be made more efficient. As described above, the data stored in the distribution shape holding unit 121 of this embodiment supports the design of input parameters in another optimization calculation in which the shape of the evaluation function is expected to be similar. Is possible. In particular, adopting the real-time display function as in this embodiment enables prompt input parameter design support when it is desired to proceed with a plurality of similar optimization calculations in parallel.

As explained in the above example, in an optimization method that searches for optimal parameters by repeatedly generating search points from a certain distribution and updating the distribution based on its evaluation value, if the distribution update is unbiased, The existence of a flat region in the evaluation function causes deterioration in search performance. According to the technology provided by this embodiment, after accepting registration of parameters to be searched and their evaluation functions, generation of search points based on a certain distribution, calculation of their evaluation values, and updating of the distribution based on the evaluation values are repeated. It is an optimization device that outputs parameter values that make the evaluation value as large as possible, and has a database that holds time-series information on distribution shapes, and uses that information to search for each parameter. By determining whether or not to perform the search and updating the hyperparameters accordingly, it is possible to improve the efficiency of the search.

1 optimization device, 11 input section, 12 optimization calculation section, 13 output section, 101 processor, 102 memory, 103 auxiliary storage device, 104 input device, 105 output device, 106 communication IF, 107 internal communication line, 121 distribution shape Holding unit, 122 Search parameter selection unit, 123 Search point generation unit, 124 Evaluation value calculation unit, 125 End determination unit, 126 Distribution shape update unit

Claims (15)

An optimization device that optimizes parameter values,
an input section that receives a parameter to be searched and an evaluation function serving as an evaluation index thereof;
an optimization calculation unit that calculates an optimal value of the parameter value based on the evaluation function;
an output unit that outputs the optimal value;
including;
The optimization calculation unit includes:
a search point generation unit that generates a search point that is a value of a parameter to be evaluated from a search point generation distribution;
an evaluation value calculation unit that calculates an evaluation value of the search point based on the evaluation function;
a distribution shape updating unit that updates the search point generation distribution based on the evaluation value;
a distribution shape holding unit that holds quantities characterizing the search point generation distribution as time series information;
a search parameter selection unit that selects parameters to search based on the time series information;
a termination determination unit that determines termination based on predetermined termination conditions;
Optimizer with The optimization device according to claim 1,
The search parameter selection unit selects parameters for which the change in the distribution shape is greater than or equal to a threshold value based on time-series information of the distribution shape of the search point generation distribution stored in the distribution shape storage unit. Optimizer. The optimization device according to claim 1,
The optimization device is characterized in that the updating method of the distribution shape updating unit is such that a change in the distribution shape becomes zero on average for the evaluation function in which there is no correlation between the evaluation value and the search point. The optimization device according to claim 1,
The optimization calculation unit calculates the optimal value of the parameter value by loop processing,
An optimization device, wherein the distribution shape holding unit holds, as time-series information, a quantity characterizing the shape of the search point generation distribution updated by the distribution shape updating unit in each loop process. The optimization device according to claim 1,

The optimization device, wherein the search parameter selection unit changes the value of the hyperparameter based on the selection result of the search parameter. The optimization device according to claim 1,
An optimization device, wherein the quantity characterizing the search point generation distribution is a variance value. The optimization device according to claim 1,
An optimization device, wherein parameters not selected by the search parameter selection unit are fixed to constant values corresponding to center values of the search point generation distribution. The optimization device according to claim 1,
The optimization device, wherein the search parameter selection unit includes a function of determining whether to restart the search for a parameter once removed from the search target. The optimization device according to claim 1,
Among the quantities characterizing the shape of the search point generation distribution, there is a quantity expressed by a second-order matrix, and the search parameter selection unit performs diagonalization processing of the matrix, conversion device. An optimization device using the optimization device according to claim 1,
Configured as an optimization device for the object recognition function of picking robots,
The information received by the input unit includes information on the target object and a group of images showing the target object,
The output unit outputs parameter values of a recognition processing function that can recognize whether or not the target object is included in the image, and if it is included, the position and orientation of the object,
An optimization device characterized in that the parameters are used for accurate recognition and grasping of a target object by a picking robot. The optimization device according to claim 1,
An optimization device, wherein the output unit has an output device capable of receiving calculation results during the search that reflect the degree of progress of the search and the shape of the evaluation function, and visually displaying the information even during the search. An optimization method that is executed by an information processing device including an input device, an output device, a processor, and a storage device, and optimizes parameter values,
a first step of receiving a parameter to be searched and an evaluation function serving as its evaluation index;
a second step of generating a search point, which is the value of the parameter to be searched, from the search point generation distribution;
a third step of calculating an evaluation value of the search point based on the evaluation function;
a fourth step of updating the search point generation distribution based on the evaluation value;
a fifth step of retaining the quantity characterizing the search point generation distribution as time series information;
a sixth step of selecting parameters to search based on the time series information;
Optimization method to perform. After the sixth step, return to the second step and repeat the loop process until the end determination condition is satisfied.
The optimization method according to claim 12. The second step, the third step, and the fourth step are performed based on a distribution prediction algorithm.
The optimization method according to claim 12. performing a seventh step of updating hyperparameters based on the number of parameters to explore;
The optimization method according to claim 12.

Images