JP6269121B2 - Information processing apparatus, evaluation function learning method, and program - Google Patents

Description

  The present invention relates to an information processing apparatus, an evaluation function learning method, and a program.

  As the computing power of information processing devices improves, attempts have been made to solve various combinatorial optimization problems using information processing devices. As an example of the combinatorial optimization problem, when placing cells as circuit components on a printed board, a preferred arrangement pattern from the viewpoint of a certain evaluation criterion (for example, short wiring between cells or small occupied area) The problem to search is mentioned. By letting the information processing apparatus solve the combinatorial optimization problem, it is possible to automatically obtain a preferable solution, and it is possible to support work that depends on human experience and intuition.

  Among combinatorial optimization problems, an efficient algorithm for obtaining an optimal solution is not known, and there is a problem that it is difficult to comprehensively study all combinations from the viewpoint of computational complexity. For such combinatorial optimization problems, solutions may be searched by heuristics. For example, when an optimization process for searching for a solution can be divided into a plurality of steps, optimization is performed by selecting an operation with a high evaluation value from among operation candidates for the next one step using an evaluation function. There is a method in which processing is advanced step by step. By using heuristics, it is possible to obtain an optimal solution or a suboptimal solution close to the optimal solution.

  An integrated circuit placement apparatus that automatically places cells in the layout design of a semiconductor integrated circuit has been proposed. This integrated circuit arrangement device divides a chip area into two areas by dividing lines (cut lines), and distributes the cells into the two areas so that the number of wirings (nets) crossing the cut lines is minimized. Further, the integrated circuit arrangement apparatus extracts a critical path having a large signal delay time, and moves a cell on the critical path and a cell connected to the net on the critical path so that the delay time becomes small.

  In addition, an element arrangement apparatus that determines an arrangement pattern of LSI (Large Scale Integration) elements has been proposed. This element arrangement device searches for a solution of an arrangement pattern using an algorithm or heuristic, and sequentially displays solution candidates obtained during the search on the screen. The element placement device allows operations to change the relative positional relationship and constraint conditions between elements for the solution candidates displayed on the screen, and continues the search by reflecting the changed relative positional relationship and constraint conditions To do.

  In addition, a cell placement method has been proposed in which cells of an integrated circuit are placed on a chip. In this cell arrangement method, the following three conditions are satisfied when a cell is distributed to two arrangement areas divided by a cut line. (1) The number of signal lines intersecting with the cut line is minimized, (2) the utilization rate of the grid lines capable of drawing the signal lines is equalized between the two arrangement areas, and (3) each arrangement The total cell area of the region does not exceed the area of the arrangement region.

JP-A-8-96013 JP-A-9-330350 Japanese Patent Laid-Open No. 10-256375

  As described above, when the information processing apparatus executes the optimization process, an appropriate evaluation function may be prepared. For example, an evaluation function including a variable corresponding to a feature amount extracted from data and a parameter (for example, a coefficient as a weight applied to the feature amount) is defined. However, there are cases where it is not easy for a human to define an appropriate evaluation function (for example, to determine an appropriate parameter value), such as when there are many variables and parameters.

  Therefore, it is conceivable to learn an evaluation function from sample data using a supervised machine learning technique. For example, “good” sample data and “bad” sample data are prepared from the viewpoint of humans, a good evaluation value is calculated from the feature value of the “good” sample data, and a bad evaluation value is calculated from the feature value of the “bad” sample data. An evaluation function as calculated is calculated. For example, the value of the parameter included in the evaluation function is calculated using a statistical method.

  However, when learning the evaluation function by such a method, since humans prepare sample data, the number of samples may be insufficient. As a result, the accuracy of the learned evaluation function is reduced, and the optimization processing is not performed, for example, a “bad” operation is selected because a good evaluation value is calculated for a “bad” operation as seen by humans. It may become efficient. On the other hand, data automatically generated by the information processing apparatus is not easily used for learning an evaluation function as it is because it is not given a “good” or “bad” judgment by a human.

  In one aspect, an object of the present invention is to provide an information processing apparatus, an evaluation function learning method, and a program that improve the learning accuracy of an evaluation function used for optimization processing.

  In one aspect, an information processing apparatus having a storage unit and a calculation unit is provided. The storage unit stores information on evaluation functions used for optimization processing. The calculation unit generates evaluated data, advances an optimization process on the evaluated data using an evaluation function, and calculates an evaluation value according to the state of the evaluated data after the optimization process is advanced. The evaluation function is updated based on the evaluated data and the evaluation value. In one aspect, an evaluation function learning method executed by a computer is provided. In one aspect, a program to be executed by a computer is provided.

  In one aspect, the learning accuracy of the evaluation function used for the optimization process is improved.

It is a figure which shows the example of information processing apparatus. It is a block diagram which shows the hardware example of a design apparatus. FIG. 10 is a first diagram illustrating an example of an LSI layout optimization problem. FIG. 10 is a second diagram illustrating an example of an LSI layout optimization problem. It is a block diagram which shows the function example of a design apparatus. It is a figure which shows the example of a production | generation of the sample used for machine learning. It is a figure which shows the example of the error function used for machine learning. It is a flowchart which shows the example of a procedure of parameter learning. It is a flowchart which shows the example of a procedure of 1 step optimization. It is a figure which shows the example of the optimization problem of a container packing.

Hereinafter, the present embodiment will be described with reference to the drawings.
[First Embodiment]
FIG. 1 is a diagram illustrating an example of an information processing apparatus.

  The information processing apparatus 10 according to the first embodiment generates an evaluation function used when solving a combination optimization problem by supervised machine learning. As an example of the combinatorial optimization problem, determining an arrangement pattern of circuit blocks on a printed board in LSI design can be cited. The optimization process using the evaluation function is implemented in, for example, CAD (Computer Aided Design) software that supports design work such as LSI design. The evaluation function learned by the information processing apparatus 10 can be used in the information processing apparatus 10 or another information processing apparatus.

  The information processing apparatus 10 includes a storage unit 11 and a calculation unit 12. The information processing apparatus 10 may be called a computer. The storage unit 11 may be a volatile storage device such as a RAM (Random Access Memory) or a non-volatile storage device such as an HDD (Hard Disk Drive). The calculation unit 12 is, for example, a processor. The processor may be a central processing unit (CPU) or a digital signal processor (DSP), and may include an integrated circuit for a specific application such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The processor may execute a program stored in a storage device such as a RAM (for example, the storage unit 11). A set of two or more processors (multiprocessor) may be called a “processor”.

  The storage unit 11 stores information indicating the evaluation function 13 (first evaluation function) used for the optimization process. The evaluation function 13 includes, for example, one or more variables corresponding to the feature amount indicating the feature of the evaluated data, and parameters other than the variable (for example, a coefficient corresponding to the weight of the feature amount). In this case, the information of the evaluation function 13 includes parameter values calculated by machine learning. By substituting the feature value of the evaluated data into the variable of the evaluation function 13, an evaluation value indicating the evaluation result of the evaluated data by the evaluation function 13 can be calculated. Examples of the feature amount of the data to be evaluated include the number of wiring satisfying a certain condition among wirings between circuit blocks, the total area of circuit blocks arranged in each sub-region on the printed board, and the like.

  The arithmetic unit 12 learns the evaluation function using the evaluated data as the sample data so that the accuracy of the evaluation function used for the optimization process is increased (for example, calculates the value of the parameter included in the evaluation function). As will be described later, the calculation unit 12 updates the evaluation function 13 (first evaluation function) indicated by the information stored in the storage unit 11 to the evaluation function 13a (second evaluation function). For example, the evaluation function 13 a is different in parameter value from the evaluation function 13. However, the evaluation function 13a may include variables and parameters that are not included in the evaluation function 13.

  In calculating the evaluation function 13 a, the calculation unit 12 generates the evaluated data 14. For example, the evaluated data 14 includes information on the constraint conditions of circuit blocks and arrangement patterns whose arrangement is to be determined. The evaluated data 14 need not be sample data created by a human, but may be automatically generated according to a predetermined algorithm or randomly. For example, the data 14 to be evaluated may be data obtained by randomly transforming sample data created by a human.

  When the evaluated data 14 is generated, the arithmetic unit 12 advances an optimization process on the evaluated data 14 using the evaluation function 13. In the optimization process, the calculation unit 12 selects a candidate having a good evaluation value of the evaluation function 13 from a plurality of candidates. When the optimization process includes a plurality of stages of partial optimization processes, the arithmetic unit 12 may execute the partial optimization processes up to the nth stage (n is an integer of 1 or more) in the middle of the optimization process. . n may be a predetermined fixed value. For example, the calculation unit 12 repeats n times hierarchically to divide the circuit board into two sub-regions so that the evaluation value is increased by dividing the printed board region into two.

  And the calculating part 12 calculates the evaluation value 15 according to the state of the to-be-evaluated data 14 after progressing an optimization process. For example, the calculation unit 12 calculates the evaluation value 15 by substituting the feature amount of the arrangement pattern when the area of the printed board is hierarchically divided n times into the variable of the evaluation function 13. In general, even if the quality of certain sample data is not clear at that time, it is often clear whether the original sample data was good or bad when the optimization process is advanced. That is, the evaluation value calculated from the feature value after the optimization process is clearly better or worse than the evaluation value calculated from the feature value before the optimization process. Often becomes. Therefore, the calculation unit 12 uses the evaluation value 15 after the optimization process is performed, instead of the evaluation value obtained directly from the generated evaluated data 14.

  When the evaluation value 15 is calculated, the calculation unit 12 updates the evaluation function 13 to the evaluation function 13a based on the generated evaluated data 14 and the evaluation value 15 after the optimization process is performed (for example, evaluation Update the value of the parameter of function 13). For example, the computing unit 12 calculates the parameter value of the evaluation function 13a by performing regression analysis on a plurality of feature value / evaluation value pairs including the feature value / evaluation value 15 pair of the evaluated data 14. At this time, the calculation unit 12 estimates the relationship between the feature amount of the original sample data before the optimization process is advanced and the evaluation value of the state after the optimization process is advanced. The evaluation value 15 used for the calculation of the evaluation function 13a can be said to be an “ideal evaluation value” because there is a high possibility that the evaluation data 14 is more clearly expressed as good or bad.

  This process may take time if the number of data 14 to be evaluated is very large. In that case, by using only a part of the automatically generated data, the data can be reduced and the processing time can be shortened. At this time, sample data having a small difference between the first evaluation value and the evaluation value 15 means that the conventional learning result is functioning well, so that even if it is input to the estimation process, the result is not greatly affected. However, if the estimation process is executed without selecting a small difference at all, the existing learning result may be destroyed. Therefore, when selecting data, increasing the probability that a value with a large difference in evaluation values will be selected, and selecting a value with a small difference with a low probability, the data necessary for correction can be acquired even if the data is reduced. It can be expected to reduce the possibility of spilling. For example, when the difference between the first evaluation value and the evaluation value 15 is 0 to 100, the selection probability is 1%, and when the difference between the evaluation values is 100 to 200, the selection probability is 10%. The probability corresponding to is determined, and data used for the estimation process may be selected using a random number.

  According to the information processing apparatus 10 of the first embodiment, when learning the evaluation function used for the optimization process, the evaluation value 15 after the optimization process is performed on the evaluated data 14 is calculated, The evaluation function 13 is updated based on the evaluation data 14 and the evaluation value 15. As a result, even if the evaluated data 14 is automatically generated without human intervention, the quality of the evaluated data 14 can be estimated, and the accuracy (e.g., the evaluation function) of the evaluation function 13 can be estimated using the evaluated data 14. 13) can be improved.

  Therefore, the shortage of sample data prepared by a human can be compensated, and the burden when calculating the evaluation function by supervised machine learning can be reduced. In addition, by improving the accuracy of the evaluation function, it is possible to prevent a good evaluation value from being calculated for a “bad” candidate from the viewpoint of humans, and to efficiently search for an optimal solution or a sub-optimal solution. Become.

[Second Embodiment]
Next, a second embodiment will be described. The design apparatus 100 according to the second embodiment supports LSI layout design. The design apparatus 100 can be realized as a computer. In this case, the design apparatus 100 executes CAD software used for LSI layout design.

FIG. 2 is a block diagram illustrating a hardware example of the design apparatus.
The design apparatus 100 includes a CPU 101, a RAM 102, an HDD 103, an image signal processing unit 104, an input signal processing unit 105, a medium reader 106, and a communication interface 107. The CPU 101 is an example of the calculation unit 12 according to the first embodiment. The RAM 102 or the HDD 103 is an example of the storage unit 11 according to the first embodiment.

  The CPU 101 is a processor including an arithmetic circuit that executes program instructions. The CPU 101 loads at least a part of the program and data stored in the HDD 103 into the RAM 102 and executes the program. The CPU 101 may include a plurality of processor cores, the design apparatus 100 may include a plurality of processors, and the processes described below may be executed in parallel using a plurality of processors or processor cores. A set of processors (multiprocessor) may be called a “processor”.

  The RAM 102 is a volatile memory that temporarily stores programs executed by the CPU 101 and data used by the CPU 101 for calculation. The design apparatus 100 may include a type of memory other than the RAM, or may include a plurality of memories.

  The HDD 103 is a nonvolatile storage device that stores software programs such as an OS (Operating System) and application software, and data. Application software includes CAD software used for LSI layout design. The design apparatus 100 may include other types of storage devices such as a flash memory and an SSD (Solid State Drive), and may include a plurality of nonvolatile storage devices.

  The image signal processing unit 104 outputs an image to the display 21 connected to the design apparatus 100 according to a command from the CPU 101. As the display 21, a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD), a plasma display (PDP), an organic electro-luminescence (OEL) display, or the like can be used. .

  The input signal processing unit 105 acquires an input signal from the input device 22 connected to the design apparatus 100 and outputs it to the CPU 101. As the input device 22, a pointing device such as a mouse, a touch panel, a touch pad, a trackball, a keyboard, a remote controller, a button switch, or the like can be used. A plurality of types of input devices may be connected to the design apparatus 100.

  The medium reader 106 is a reading device that reads a program and data recorded on the recording medium 23. The program may include a CAD software program used for LSI layout design. Examples of the recording medium 23 include a magnetic disk such as a flexible disk (FD) and an HDD, an optical disk such as a CD (Compact Disc) and a DVD (Digital Versatile Disc), a magneto-optical disk (MO), A semiconductor memory or the like can be used. For example, the medium reader 106 stores the program and data read from the recording medium 23 in the RAM 102 or the HDD 103.

  The communication interface 107 is an interface that is connected to the network 24 and communicates with other information processing apparatuses via the network 24. The communication interface 107 may be a wired communication interface connected to a communication device such as a switch via a cable, or may be a wireless communication interface connected to a base station via a wireless link.

  Note that the design apparatus 100 may not include at least one of the medium reader 106 and the communication interface 107. In addition, the design apparatus 100 may not include at least one of the image signal processing unit 104 and the input signal processing unit 105 when remotely accessed from a terminal device operated by a user. In addition, at least one of the display 21 and the input device 22 may be formed integrally with the housing of the design apparatus 100.

Next, an example of an LSI layout optimization problem solved using the design apparatus 100 will be described.
FIG. 3 is a first diagram illustrating an example of an LSI layout optimization problem.
Here, it is considered to search for an optimal arrangement pattern of cells as circuit components by the partitioning base arrangement method. As input data, design data indicating a plurality of cells, a net that connects the cells, and an area in which the cells can be arranged is given. Each cell is provided with one or more pins, and a pin for connecting to an external LSI may be provided at the edge of the entire region. One net is connected to two or more pins. In the example of FIG. 3, a region R in which cells # 1 to # 8 and cells # 1 to # 8 can be arranged is defined. In FIG. 3, the description of the net is omitted or simplified.

  In the partitioning-based arrangement method, the design apparatus 100 hierarchically repeats dividing the area into two and allocating a plurality of cells to the two areas until the positions of the cells are sufficiently limited. The operation of dividing a region into two and allocating cells corresponds to one step in heuristics. Since there are a plurality of methods (division patterns) for allocating cells to two areas, the design apparatus 100 selects a division pattern having the best evaluation value calculated by the evaluation function from a plurality of division patterns for each step. To do. Since this evaluation value is calculated during the optimization process, it can also be referred to as an “intermediate evaluation value”. The design apparatus 100 further divides the divided area into two as the next step based on the selected division pattern.

  As an example, the design apparatus 100 divides the region R into regions R0 and R1 having the same area by drawing a dividing line (cut line) in the X-axis direction of the region R. The design apparatus 100 generates a division pattern by arranging the cells # 1 to # 8 in any of the regions R0 and R1. For each divided pattern, the evaluation value is calculated by substituting the feature amount into the evaluation function. In the example of FIG. 3, cells # 1, # 3, # 4, and # 8 are arranged in the region R0, and cells # 2, # 5, # 6, and # 7 are arranged in the region R1. When there is a net connecting the cell arranged in the region R0 and the cell arranged in the region R1, the net intersects with the cut line. The number of nets that intersect the cut line may be referred to as the cut number.

  When the division pattern having the best evaluation value for the division of the region R is selected, the design apparatus 100 draws a cut line (perpendicular to the previous cut line) in the Y-axis direction of the region R0. R0 is divided into a region R00 and a region R01 having the same area. The design apparatus 100 generates the division pattern by arranging the cells # 1, # 3, # 4, and # 8 arranged in the region R0 in the previous step in any of the regions R00 and R01. In the example of FIG. 3, cells # 1 and # 4 are arranged in the region R00, and cells # 3 and # 8 are arranged in the region R01.

  Similarly, the design apparatus 100 divides the region R1 into regions R10 and R11 having the same area by drawing a cut line in the Y-axis direction of the region R1. The design apparatus 100 generates a division pattern by arranging the cells # 2, # 5, # 6, and # 7 arranged in the region R1 in the previous step in any of the regions R10 and R11. In the example of FIG. 3, cells # 5 and # 6 are arranged in the region R10, and cells # 2 and # 7 are arranged in the region R11. As described above, in the partitioning-based arrangement method, the region is subdivided based on the selected upper division pattern and a plurality of partial optimization problems are generated hierarchically.

  Here, various criteria can be considered as criteria for selecting a “good” division pattern from a plurality of division patterns for each step. As an example, there are a mini-cut method, a balance-considering cut method, a weighted mini-cut method, and the like as focusing on the number of cuts.

  In the mini-cut method, the number of cuts, that is, the number of nets intersecting the cut line is used as an evaluation value, and a division pattern having a small evaluation value is defined as a good division pattern. However, the constraint is that the total area of cells arranged in one region is equal to or substantially equal to the total area of cells arranged in the other region (for example, the difference between the two is equal to or less than a threshold).

The balance-considering cut method allows a division pattern in which the total area of cells is not balanced between two regions to a certain extent. In the balance-considering cut method, for example, an evaluation function is given as in Expression (1). Cuts represents the number of cuts, D ₀ represents the total area of cells arranged in one region, and D ₁ represents the total area of cells arranged in the other region. The second term of the formula (1) is a term indicating the deviation of the total area of the cells. w _a is a weight given to the number of cuts, and w _b is a weight given to the deviation of the total area. That is, two of the number of cuts and the deviation of the total area are evaluated based on a certain weight. When the weights w _a and w _b are changed, the evaluation of the quality of each divided pattern can also be changed. As will be described later, the weights w _a and w _b can be learned by supervised machine learning performed before searching for an optimum LSI layout.

The weighted mini-cut method does not evaluate all nets equally but weights and evaluates nets according to the number of pins connected to the nets. In the weighted mini-cut method, nets that intersect the cut line are classified according to the number of pins, and the number of cuts is counted for each number of pins. In the weighted mini-cut method, for example, an evaluation function is given as in Expression (2). Cuts (i) is the number of cuts for a net with pin number i, and w _i is a weight given to Cuts (i). The evaluation value is the sum of w ₂ × Cuts (2), w ₃ × Cuts (3),... The smaller the evaluation value, the better the division pattern. The weight w _i can be learned by supervised machine learning as will be described later.

FIG. 4 is a second diagram illustrating an example of an LSI layout optimization problem.
Examples of other optimization problems in which blocks as circuit components are arranged on a printed board by a partitioning base arrangement method will be given. Here, it is assumed that each net is connected to two pins and connects two blocks. A net group obtained by grouping one or more nets is connected between the two blocks. Some net groups are given a maximum line length constraint that the length of the wiring is X or less. In this optimization problem, attention is paid to the area required to satisfy the maximum line length constraint.

  For example, the evaluation function is given as in Expression (3). D represents the total area of blocks arranged in a certain region. Both of the blocks at both ends are included in the area of the area s, and for a net group including n nets without a maximum line length constraint, Na0 (n) is the number of the net groups, and Wa0 (n) is the corresponding The use width of the net group, La0 (s, n), represents an estimate of the line length of the net group. If a block is arranged in a narrow region, the wiring may become long, so the net group line length depends on the region area s.

  Also, both of the blocks at both ends are included in the area of the area s, and for a net group that has a maximum line length constraint and includes n nets, Na1 (n) is the number of the net groups, Wa1 (n) Is the use width of the net group, and La1 (s, n) is an estimate of the line length of the net group. Further, each of the blocks at both ends is included in the region of the area s, and another net group having a maximum line length restriction is connected to at least one block, and a net group including n nets is Na2 (N) represents the number of the net groups, Wa2 (n) represents the usage width of the net group, and La2 (s, n) represents an estimate of the line length of the net group. This is because not only the block connected to the net group with the maximum line length restriction but also the adjacent block is subject to the arrangement restriction. One net group may be counted as Na0 or both Na1 and Na2.

  Further, only one of the blocks at both ends is included in the area of the area s, and there is no maximum line length restriction, and for a net group including n nets, Nb0 (n) is the number of the net groups, and Wb0 (n ) Represents the usage width of the net group, and Lb0 (s, n) represents an estimate of the line length of the net group. Further, only one of the blocks at both ends is included in the region of the area s, and for a net group that includes a maximum line length constraint and includes n nets, Nb1 (n) is the number of the net groups, and Wb1 (n ) Represents the usage width of the net group, and Lb1 (s, n) represents an estimate of the line length of the net group. In addition, only one of the blocks at both ends is included in the area of the area s, and another net group having a maximum line length restriction is connected to at least one block, and the net group includes n nets. Nb2 (n) represents the number of the net group, Wb2 (n) represents the use width of the net group, and Lb2 (s, n) represents an estimate of the line length of the net group.

  Also, none of the blocks at both ends is included in the region of the area s, and for a net group including n nets without a maximum line length constraint, Nc0 (n) is the number of the net groups, Wc0 (n ) Represents the used width of the net group, and Lc0 (s, n) represents an estimate of the line length of the net group. Also, none of the blocks at both ends is included in the area of the area s, and for a net group that includes the maximum number of lines and includes n nets, Nc1 (n) is the number of the net groups, and Wc1 (n ) Represents the use width of the net group, and Lc1 (s, n) represents an estimate of the line length of the net group. Also, none of the blocks at both ends is included in the region of the area s, and another net group having a maximum line length restriction is connected to at least one block, and the net group includes n nets. Nc2 (n) represents the number of the net group, Wc2 (n) represents the use width of the net group, and Lc2 (s, n) represents an estimate of the line length of the net group.

  Na0 (n), Na1 (n), Na2 (n), Nb0 (n), Nb1 (n), Nb2 (n), Nc0 (n), Nc1 (n), Nc2 (n) are extracted from the division pattern This is a variable indicating the feature amount to be performed. On the other hand, Wa0 (n), La0 (s, n), Wa1 (n), La1 (s, n), Wa2 (n), La2 (s, n), Wb0 (n), Lb0 (s, n), Wb1 (n), Lb1 (s, n), Wb2 (n), Lb2 (s, n), Wc0 (n), Lc0 (s, n), Wc1 (n), Lc1 (s, n), Wc2 ( n) and Lc2 (s, n) are parameters and can be learned by supervised machine learning as will be described later.

  For example, the design apparatus 100 calculates a value based on Expression (3) for each of the two divided areas, and uses the sum of the values of the two areas as the evaluation value of the divided pattern. The evaluation value calculated from the evaluation formula of Expression (3) represents an estimated area including the total area of the blocks and the total area of the net. A small evaluation value can be defined as a good division pattern.

  In the example of FIG. 4, a certain region is divided by a cut line into a region R0 having a region area s and a region R1 having a region area s. As a certain division pattern, blocks # 1, # 2, # 3, and # 4 are arranged in the region R0, and block # 5 is arranged in the region R1.

  A net group including 16 nets with no maximum line length restriction is connected between the block # 1 and the block # 2 in the region R0. Therefore, the parameters Wa0 (16) and La0 (s, 16) are applied to this net group. In the region R0, a net group including 16 nets with a maximum line length restriction is connected between the block # 3 and the block # 4. Therefore, the parameters Wa1 (16) and La1 (s, 16) are applied to this net group. A net group including 12 nets without a maximum line length constraint is connected between the block # 2 and the block # 3 in the region R0. Further, as described above, another net group having a maximum line length constraint is connected to the block # 3. Therefore, the parameters Wa0 (12), La0 (s, 12) and the parameters Wa2 (12), La2 (s, 12) are applied to this net group.

  A net group including six nets having no maximum line length constraint extends from block # 1 in the region R0 toward another region. Therefore, the parameters Wb0 (6) and Lb0 (s, 6) are applied to this net group when calculating the evaluation value of the region R0. A net group including eight nets with no maximum line length constraint extends from block # 2 in the region R0 toward another region. Therefore, the parameters Wb0 (8) and Lb0 (s, 8) are applied to this net group in calculating the evaluation value of the region R0.

  A net group including eight nets having no maximum line length constraint extends from block # 5 in the region R1 toward another region. Therefore, the parameters Wb0 (8) and Lb0 (s, 8) are applied to this net group in calculating the evaluation value of the region R1. Further, a net group including six nets that are not located in the region R1 and have no maximum line length constraint passes through the region R1. Therefore, the parameters Wc0 (6) and Lc0 (s, 6) are applied to this net group in calculating the evaluation value of the region R1.

Next, machine learning of evaluation function parameters in the design apparatus 100 will be described.
FIG. 5 is a block diagram illustrating an example of functions of the design apparatus.
The design apparatus 100 includes a manual sample storage unit 111, an automatic sample storage unit 112, a parameter storage unit 113, an initial value calculation unit 114, an automatic sample generation unit 115, a regression analysis unit 117, a parameter update unit 118, and a variable setting unit 119. . The manual sample storage unit 111, the automatic sample storage unit 112, and the parameter storage unit 113 are mounted as storage areas secured in the RAM 102 or the HDD 103, for example. The initial value calculation unit 114, the automatic sample generation unit 115, the regression analysis unit 117, the parameter update unit 118, and the variable setting unit 119 are implemented as, for example, modules of a program executed by the CPU 101.

  The manual sample storage unit 111 stores a sample (manual sample) created by a human, which is a division pattern sample used for parameter learning. A manual sample can be used as an example of a good division pattern. The manual sample includes, for example, information such as a plurality of blocks, a plurality of nets connecting the blocks, a region area, and a position of each block.

  The automatic sample storage unit 112 stores a sample (automatic sample) of a division pattern used for parameter learning and automatically generated by the automatic sample generation unit 115. An automatic sample may be a good division pattern or a bad division pattern. The automatic sample includes, for example, information such as a plurality of blocks, a plurality of nets connecting between the blocks, a region area, and a distribution result of each block. Further, an “ideal evaluation value” calculated by the automatic sample generation unit 115 is given to each automatic sample. It can be said that the ideal evaluation value represents an estimation result of the quality of the automatic sample.

  The parameter storage unit 113 stores parameter information including parameters included in the evaluation function and calculated parameter values. The initial value calculation unit 114 calculates the initial value of the parameter and stores it in the parameter storage unit 113. Thereafter, the parameter update unit 118 may update the parameter value once or twice or more so as to increase the accuracy. In the case of the example of FIG. 4 and Formula (3), parameters such as Wa0 (1), La0 (s, 1), Wa0 (2), La0 (s, 2), Wa0 (3), La0 (s, 3) Is stored in the parameter storage unit 113. The parameter value may be prepared by limiting the number of nets n to an integer range of 1 or more and a predetermined maximum value or less. For a net group including a number of nets exceeding the maximum value, n = maximum value. The value of the parameter corresponding to may be applied.

  The initial value calculation unit 114 performs supervised machine learning using the manual samples stored in the manual sample storage unit 111 as teacher data, and calculates initial values of parameters of the evaluation function. First, the initial value calculation unit 114 sets a manual sample as a good sample indicating a good division pattern. Further, the initial value calculation unit 114 automatically generates a random division pattern and sets a sample different from the manual sample as a bad sample indicating a bad division pattern.

  Then, the initial value calculation unit 114 calculates the evaluation function parameters so that a good evaluation value (for example, a low evaluation value) is calculated from a good sample and a bad evaluation value (for example, a high evaluation value) is calculated from a bad sample. To learn. For the parameter learning, for example, a machine learning method such as SVM (Support Vector Machine) is used. In SVM, a boundary surface that can most clearly divide a good sample and a bad sample in the vector space, that is, a boundary surface that maximizes both the distance to the good sample and the distance to the bad sample is calculated. The initial value calculation unit 114 stores the calculated parameter value in the parameter storage unit 113.

  The automatic sample generation unit 115 generates an automatic sample different from a manual sample and calculates an ideal evaluation value that estimates the quality of the automatic sample in order to increase the accuracy of the parameter value. First, the automatic sample generation unit 115 generates an automatic sample different from the manual sample through, for example, deformation of the manual sample stored in the manual sample storage unit 111.

  Next, the automatic sample generation unit 115 starts the generated automatic sample and advances the optimization process by a predetermined step (for example, 3 steps). At this time, as the parameter value of the evaluation function used for selecting the division pattern at each step, the value currently stored in the parameter storage unit 113 (for example, the initial value calculated by the initial value calculation unit 114) is used. For example, the automatic sample generation unit 115 generates a division pattern in which each region indicated by the division pattern of the automatic sample is further divided into eight by performing the optimization process three steps. The automatic sample generation unit 115 includes a one-step optimization unit 116 that advances the optimization process step by step.

  Then, the automatic sample generation unit 115 calculates an ideal evaluation value corresponding to the division pattern after the optimization process is advanced by a predetermined step. The ideal evaluation value can be calculated by substituting the feature value of the divided pattern after the optimization process is advanced by a predetermined step into an evaluation function to which the parameter value stored in the parameter storage unit 113 is applied. For example, the sum of the evaluation values of each area divided into eight is set as an ideal evaluation value corresponding to the automatic sample. In other words, instead of the evaluation value calculated from the feature value of the automatic sample division pattern itself, the evaluation value calculated from the feature value of the division pattern after the optimization process has been advanced by a predetermined step is used for parameter learning. Will be. Since it is highly likely that the quality of the division pattern is clearer than the time of the automatic sample by proceeding with the optimization process, it can be said that the ideal evaluation value is an estimate of the quality of the automatic sample. The automatic sample generation unit 115 associates the generated automatic sample with the calculated ideal evaluation value and stores them in the automatic sample storage unit 112. In this process, when the number of automatic samples is very large, the automatic sample generation unit 115 reduces the data by using only a part of the data, as in the first embodiment, and performs processing. Time can be reduced. At this time, the selection is performed probabilistically based on random numbers, and the probability of storing a sample having a large difference between the ideal evaluation value and the current evaluation value in the automatic sample storage unit 112 is increased, so that correction can be made even if data is reduced. It can be expected to reduce the possibility of missing necessary data.

  The regression analysis unit 117 performs regression analysis using the automatic sample and ideal evaluation value stored in the automatic sample storage unit 112 to calculate the parameter value. First, the regression analysis unit 117 extracts feature amounts such as the total area of blocks and the number of net groups (D in Equation (3), Na0 (1), Na0 (2), etc.) from the division pattern of each automatic sample. . Then, the regression analysis unit 117 minimizes the error between the evaluation value calculated by substituting the feature amount extracted from the automatic sample into the evaluation function and the ideal evaluation value associated with the automatic sample. Adjust the parameter value. For parameter learning, for example, a regression analysis method such as ε-SVR (Support Vector Regression) can be used.

  The parameter update unit 118 determines whether the parameter value calculated by the regression analysis unit 117 is improved as compared with the current parameter value stored in the parameter storage unit 113. If it is determined that the parameter has been improved, the parameter updating unit 118 updates the parameter value stored in the parameter storage unit 113 to the parameter value calculated by the regression analysis unit 117. In that case, the design apparatus 100 may proceed with learning of the parameter again using the updated parameter value. On the other hand, when it is determined that there is no improvement, the parameter update unit 118 does not employ the parameter value calculated by the regression analysis unit 117. In this case, the design apparatus 100 may end the parameter learning.

  Whether the parameter value is improved can be verified using a sample such as an automatic sample. For example, the design apparatus 100 performs an optimization process from a division pattern of a certain sample to the end (until the position of each block is sufficiently limited) to generate a final division pattern. The parameter update unit 118 confirms the correlation between the evaluation value of the initial division pattern calculated using the current parameter value and the quality of the final division pattern. Further, the parameter update unit 118 confirms the correlation between the evaluation value of the initial division pattern calculated using the new parameter value and the quality of the final division pattern.

  Then, the parameter updating unit 118 determines that the parameter value is improved when the correlation is improved. A good evaluation value is calculated for the original division pattern when the division pattern after optimization is “good”, and a bad evaluation value is calculated for the original division pattern when the division pattern after optimization is “bad”. If so, it can be said that the correlation is strong. The parameter update unit 118 may calculate a correlation coefficient as an index value indicating the strength of correlation.

  The variable setting unit 119 allows a new variable or a new parameter to be added to the evaluation function in response to an input from the user. When the user gives an instruction to add a variable or parameter, the variable setting unit 119 notifies the regression analysis unit 117 of the variable or parameter to be added. Then, the next time regression analysis is performed, the regression analysis unit 117 extracts a new feature amount corresponding to the added variable from the automatic sample, and performs regression analysis on the new evaluation function to which the variable and the parameter are added. The addition of variables and parameters can be performed while the automatic sample generation unit 115, the regression analysis unit 117, and the parameter update unit 118 are repeatedly learning the parameters. Therefore, the user can appropriately add preferable variables and parameters that improve the evaluation value to the evaluation function while watching the learning status of the evaluation function.

FIG. 6 is a diagram illustrating a generation example of a sample used for machine learning.
As described above, the automatic sample generation unit 115 generates an automatic sample based on a manual sample. This can cover the shortage of manpower samples. As an automatic sample generation method, various generation methods can be considered according to the contents of the optimization problem.

  For example, the automatic sample generation unit 115 randomly extracts a partial area from the LSI layout indicated by the manual sample and adopts it as a partial problem. That is, the automatic sample generation unit 115 initializes the layout of the cut-out area by erasing the block position information while maintaining the information on the blocks included in the cut-out area and the nets connecting the blocks. Use as a base for At this time, the net that connects the inside and the outside of the cut-out area is handled on the assumption that a pin is set at the edge of the area. The automatic sample generation unit 115 uses, as an automatic sample, a division pattern in which optimization processing has been advanced by one step for a region that has been cut out and initialized (the region has been divided once).

  In the example of FIG. 6, blocks # 1 to # 8 are arranged in the entire area indicated by the manual sample. Block # 1 and block # 6, block # 2 and block # 7, and block # 5 and block # 8 are connected via the net. It is assumed that the automatic sample generation unit 115 extracts an area including blocks # 1 to # 5 but not including blocks # 6 to # 8 from the entire area. Then, the automatic sample generation unit 115 deletes the position information of the blocks # 1 to # 5. In addition, the automatic sample generation unit 115 sets a pin at a point where the net between the blocks # 1 and # 6 and the edge of the region intersect, and treats the pin and the block # 1 as being connected by the net. The net between blocks # 2 and # 7 and the net between blocks # 5 and # 8 are handled in the same manner. Then, the automatic sample generation unit 115 generates a division pattern in which the area is divided into two and blocks # 1 to # 5 are allocated as an automatic sample.

As another automatic sample generation method, a method of randomly changing the design conditions such as the block size, the number of wiring layers, and the wiring pits from that of a manual sample can be considered.
FIG. 7 is a diagram illustrating an example of an error function used for machine learning.

  As described above, the initial value calculation unit 114 assigns a good evaluation value to a good sample and a bad evaluation value to a bad sample from the feature amount extracted from the sample of the division pattern and the determination result of the quality of the sample. To learn the parameters of the evaluation function. For example, the initial value calculation unit 114 searches for a parameter value that minimizes the value of Expression (4). In the search for the parameter value, for example, a search method such as a steepest descent method can be used.

When using equation (4), all combinations of good and bad samples are listed. diff _j represents the difference between the evaluation value of the good sample and the evaluation value of the bad sample for the j-th combination. err0 (x) is an error function, and L is a predetermined positive value. As shown in FIG. 7A, err0 (x) returns 0 when x is −L or less, and returns x + L when x is larger than −L. That is, the first term of the equation (4) indicates the purpose of minimizing the sum of the values of diff _j greater than -L. λ is a predetermined value, and w _i represents the i-th parameter included in the evaluation function. The second term of Equation (4) is a normalization term that suppresses overlearning. However, the sum of squares of parameter values may be replaced with the sum of absolute values of parameter values.

  Further, as described above, the regression analysis unit 117 can most appropriately express the relationship between the feature value and the ideal evaluation value from the feature value extracted from the sample of the divided pattern and the ideal evaluation value of the sample. Learn the parameters of the evaluation function. For example, the regression analysis unit 117 searches for parameter values that minimize the value of Equation (5). In the search for the parameter value, for example, a search method such as a steepest descent method can be used.

In Expression (5), v _j represents an ideal evaluation value of the j-th sample, and a _j represents a set of feature quantities of the j-th sample. Therefore, v _j −f (a _j ) represents an error between the ideal evaluation value of the j-th sample and the evaluation value when the feature amount of the j-th sample is substituted into the evaluation function. err1 (x) is an error function, and ε is a predetermined positive value. As shown in FIG. 7B, err1 (x) returns 0 when the absolute value of x is equal to or smaller than ε, and returns | x | −ε when the absolute value of x is larger than ε. That is, the first term of Equation (5) indicates the purpose of minimizing the sum of the absolute values of the evaluation value errors exceeding ε. However, the absolute value of the evaluation value error may be replaced with the square of the evaluation value error. Similar to Equation (4), the second term is a normalization term.

Next, a machine learning procedure of parameters executed by the design apparatus 100 will be described.
FIG. 8 is a flowchart illustrating an exemplary procedure for parameter learning.
(S10) The initial value calculation unit 114 acquires a plurality of division pattern samples (manual samples) stored in the manual sample storage unit 111 as good samples.

(S11) The initial value calculation unit 114 randomly generates a plurality of division patterns, and adopts a generated division pattern that is different from a manual sample as a bad sample.
(S12) The initial value calculation unit 114 calculates the initial value of the parameter of the evaluation function using the good sample in step S10 and the bad sample in step S11. For example, the initial value calculation unit 114 extracts feature quantities such as the number of net groups Na0 (1), Na0 (2),... From each of a good sample and a bad sample. The initial value calculation unit 114 uses a supervised machine learning method such as SVM so that a good evaluation value is calculated from a good sample feature amount and a bad evaluation value is calculated from a bad sample feature amount. The values of parameters such as Wa0 (1), Wa0 (2),... And line lengths La0 (1), La0 (2),. The initial value calculation unit 114 stores the calculated parameter value in the parameter storage unit 113.

  (S13) The automatic sample generation unit 115 randomly changes a block or net that appears by cutting out a partial area from the manual sample stored in the manual sample storage unit 111 and deleting the position information of each block. Design data to be generated automatically. However, the automatic sample generation unit 115 may automatically generate design data by other methods such as randomly changing design conditions such as the block size, the number of wiring layers, and wiring pits. The automatic sample generation unit 115 may select a generation method according to the content of the optimization problem.

  (S14) The automatic sample generation unit 115 advances the optimization process by one step by calling the one-step optimization unit 116 for the design data for the region cut out in step S13. At this time, an evaluation function to which the initial value of the parameter calculated in step S12 (or the parameter value updated in step S20 below) is applied is used. For example, the one-step optimizing unit 116 divides the area cut out in step S13 into two, and allocates blocks to two areas. The automatic sample generation unit 115 acquires the division pattern after one step as an automatic sample P.

  (S15) The automatic sample generation unit 115 further advances the optimization process by three steps by calling the one-step optimization unit 116 a plurality of times for the automatic sample P generated in step S115. At this time, an evaluation function to which the initial value of the parameter calculated in step S12 (or the parameter value updated in step S20 below) is applied is used. For example, each of the regions divided into two in step S14 is further divided into eight.

  (S16) The automatic sample generation unit 115 calculates an ideal evaluation value from the division pattern after step S15 is executed. For example, for each subdivided area, the automatic sample generation unit 115 uses the feature amounts such as the number of net groups Na0 (1), Na0 (2),... And the usage widths Wa0 (1), Wa0 (2),. An evaluation value for the estimated area calculated from parameter values such as line lengths La0 (1), La0 (2),. The automatic sample generation unit 115 sets the sum of the evaluation values of the subdivided regions (the total of the estimated areas) as the ideal evaluation value. However, the method of summing up the evaluation values of each region is an example, and the ideal evaluation value may be calculated by another method. The automatic sample generation unit 115 stores the calculated ideal evaluation value and the automatic sample P generated in step S14 in the automatic sample storage unit 112 in association with each other.

  (S17) The automatic sample generation unit 115 determines whether a sufficient number of automatic samples P are obtained to perform the regression analysis. The required number of automatic samples P may be designated in advance by the user. If the required number of automatic samples P is obtained, the process proceeds to step S18. If the required number of automatic samples P is not obtained, the process proceeds to step S13.

  (S18) The regression analysis unit 117 calculates the value of the parameter of the evaluation function using a plurality of combinations of the automatic sample P and the ideal evaluation value. For example, the regression analysis unit 117 obtains a feature amount from each automatic sample P (before three steps are advanced in step S15) such as the number of net groups Na0 (1), Na0 (2),. Extract. Then, the regression analysis unit 117 uses a regression analysis method such as ε-SVR so that the error between the evaluation value calculated from the evaluation function and the ideal evaluation value is minimized. The values of parameters such as Wa0 (2),..., Line length La0 (1), La0 (2),.

  (S19) The parameter update unit 118 verifies whether the parameter value calculated in step S18 is improved from the current parameter value stored in the parameter storage unit 113. If it is determined that the parameter value has improved, the process proceeds to step S20. If it is determined that the parameter value has not improved, the process proceeds to step S21.

  In verifying the improvement, for example, the design apparatus 100 proceeds with optimization processing to the end of each of the plurality of samples to determine whether the final block arrangement is acceptable. The parameter update unit 118 determines the correlation between the evaluation value of the initial division pattern calculated using the current parameter value and the quality of the final block arrangement (for example, calculates an index value such as a correlation coefficient). To do). Further, the parameter update unit 118 determines the correlation between the evaluation value of the initial division pattern calculated using the new parameter value and the quality of the final block arrangement. If the latter correlation is improved from the former, the parameter update unit 118 can determine that the parameter value has improved. However, the presence or absence of improvement may be determined by other methods.

(S20) The parameter update unit 118 updates the parameter value stored in the parameter storage unit 113 to the value calculated in step S18.
(S21) The automatic sample generation unit 115 determines whether the parameter learning end condition is satisfied. Examples of the termination condition include that the parameter value is not improved, and that the processes in steps S13 to S20 are repeated a predetermined number of times. If the end condition is not satisfied, the process proceeds to step S13, and parameter learning is continued. If the termination condition is satisfied, the parameter learning is terminated.

FIG. 9 is a flowchart illustrating an example of a procedure for one-step optimization.
The processing of this flowchart is executed in steps S14 and S15 described above.
(S30) The 1-step optimization unit 116 divides the region into two.

  (S31) The one-step optimizing unit 116 generates one division pattern by allocating a plurality of blocks included in the region before division to two divided regions. As a method for searching for the division pattern, for example, an approximation algorithm such as a greedy method or a local search method is used.

  (S32) The one-step optimizing unit 116 extracts a feature amount from the division pattern generated in step S31. The one-step optimization unit 116 substitutes the extracted feature quantity into the evaluation function to which the current parameter value stored in the parameter storage unit 113 is applied, thereby evaluating the evaluation value indicating the quality of the divided pattern in step S31. Is calculated. Since this evaluation value indicates an evaluation in the middle of the optimization process, it can also be called an “intermediate evaluation value”.

  (S33) The one-step optimizing unit 116 determines whether or not the evaluation value calculated in step S32 is better than the currently recorded value (the best evaluation value calculated in the past). When a smaller evaluation value indicates a better evaluation, the one-step optimization unit 116 determines whether the evaluation value calculated in step S32 is smaller than that currently recorded. If the evaluation value has improved, the process proceeds to step S34. If the evaluation value has not improved (if the evaluation value is the same or worse than the currently recorded value), the process proceeds to step S35. When there is no evaluation value recorded (when it is the evaluation value of the first division pattern), it is treated as improved.

  (S34) The 1-step optimizing unit 116 records the division pattern generated in step S31 as the best division pattern searched so far. The evaluation value calculated in step S32 is recorded as the best evaluation value in the range searched up to now. At this time, the previously recorded division pattern and evaluation value may be deleted.

  (S35) The one-step optimizing unit 116 determines whether or not the division pattern search end condition is satisfied. The termination condition depends on the approximation algorithm used in step S31. Examples of the termination condition include that a predetermined number of division patterns have been generated, and that the evaluation value has not improved continuously for a predetermined number of times. If the end condition is satisfied, the process proceeds to step S36. If the end condition is not satisfied, the process proceeds to step S31.

  (S36) The one-step optimizing unit 116 acquires the last recorded division pattern, that is, the division pattern having the best evaluation value (for example, the smallest evaluation value) among the searched ones. Output as stepped division pattern.

  The learning of the evaluation function by the design apparatus 100 has been described above using the LSI layout optimization problem as an example. However, the evaluation function learning method of the second embodiment can also be applied to optimization problems in other fields other than LSI layout. Below, the optimization problem of a container packing is demonstrated as an example of another optimization problem.

FIG. 10 is a diagram illustrating an example of a container packing optimization problem.
Here, it is considered that a plurality of rectangular parallelepiped cargoes having various sizes are distributed and packed in a plurality of containers of the same size. Among various packing patterns, the one that uses fewer containers is evaluated as a good packing pattern.

  For example, feature quantities as shown in FIG. 10 are extracted from a certain packing pattern. The m containers to be used are sorted based on the filling rate (ratio of the total capacity of the cargo to the capacity of the container), and an identification number is assigned to each container in descending order of filling rate. In addition, cargo is classified into a plurality of categories according to its size. For example, the unit length of the width (x axis) is Ux, the unit length of the depth (y axis) is Uy, and the unit length of the height (z axis) is Uz. Cargo that falls within (Ux, Uy, Uz) exceeds the size of Category # 1, Division # 1, but cargo that falls within (2Ux, Uy, Uz) exceeds the size of Division # 2, Division # 2, but (2Ux , 2Uy, Uz) Define multiple categories such that cargo that falls within category # 3 and cargo that exceeds the size of category # 2 but falls within (3Ux, Uy, Uz) are category # 4 Can do.

Then, the number of cargoes in each section can be extracted as a feature amount for each container. In FIG. 10, N _{i, j} represents the number of cargoes in section #i packed in the jth container (container #j). In container packing, for example, an evaluation function is given as shown in Equation (6). Fullness represents the fullness rate of the entire used container. w ₀ represents the weight given to the satisfaction rate, and w _{i, j} represents the weight given to N _{i, j} . In this evaluation function, Fullness, N _{i, j} are variables, and w ₀ , w _{i, j} are parameters.

  For this optimization problem, for example, a tabu search method may be used as a search algorithm. For example, Zhihong Jin, Takahiro Ito and Katsuhisa Ohno, "The Three-Dimensional Bin Packing Problem and Its Practical Algorithm", JSME (Japan Society of Mechanical Engineers) International Journal, Vol. 46 No. 1, pp.60-66, 2004-06-25.

  When the tabu search method is used, first, cargo is randomly allocated to a plurality of containers. In each container, the position of each cargo in the container can be calculated by recursively using the tabu search method (for example, by dividing an area in the container). Then, starting with an initial packing pattern, several cargoes are moved between containers. For example, one or more cargoes contained in a certain container are moved to an empty container or another container that is not empty. Further, for example, one or two or more cargoes contained in a container and one or two or more cargoes contained in another container are exchanged.

  One movement of several cargoes corresponds to one step in heuristics. The one having the best evaluation value calculated by the evaluation function is selected from a plurality of packing patterns obtained by the movement, and the next step is executed based on the selected packing pattern. At this time, the packing pattern adopted most recently N times (for example, the latest seven times) is stored as a tabu list, and the packing pattern included in the tabu list is not used in the next step. This prevents repeated movement of the same cargo.

Similar to the LSI layout optimization, the design apparatus 100 can learn the parameters w ₀ , w _{i, j} of the evaluation function used for the container packing optimization by supervised machine learning. For example, the design apparatus 100 acquires a packing pattern created by a human as a manual sample, and generates a packing pattern in which several cargoes are randomly exchanged from the manual sample. Then, the design apparatus 100 evaluates the parameter of the evaluation function so that the movement of the cargo returned to the original manual sample from the generated packing pattern is a good exchange operation and the movement of other cargo is a bad exchange operation. The initial value of is calculated.

  In addition, the design apparatus 100 generates an automatic sample of the packing pattern by, for example, moving a cargo at random several steps from a manual sample. Then, the design apparatus 100 uses the evaluation function to which the initial value of the parameter calculated above is applied to advance the optimization process for several steps (for example, three steps), and the packing after the several steps are advanced. An ideal evaluation value is calculated from the pattern. The design apparatus 100 can perform regression analysis on the generated automatic sample and the calculated ideal evaluation value, and update the parameter value.

  According to the design apparatus 100 of the second embodiment, when learning the evaluation function used for the optimization process, an automatic sample is generated, and the evaluation value after the optimization process is performed on the automatic sample is an ideal evaluation. Calculated as a value. The parameter value of the evaluation function is updated by performing regression analysis on the automatic sample and the ideal evaluation value. Thereby, the quality of the automatic sample can be estimated, and the accuracy of the parameter of the evaluation function can be improved using the automatic sample. Therefore, the shortage of human samples prepared by humans can be compensated, and the burden when calculating the evaluation function by supervised machine learning can be reduced. In particular, even when there are a large number of variables and parameters, a sufficient number of samples can be prepared for regression analysis.

  In addition, by improving the accuracy of the evaluation function, it is possible to prevent a good evaluation value from being calculated for a “bad” candidate from the viewpoint of humans, and to efficiently search for an optimal solution or a sub-optimal solution. Become. In addition, it is possible to appropriately modify the evaluation function by the user, for example, by adding preferable new variables and parameters while repeating learning of the evaluation function.

  As described above, the information processing according to the first embodiment can be realized by causing the information processing apparatus 10 to execute a program. The information processing of the second embodiment can be realized by causing the design apparatus 100 to execute a program.

  The program can be recorded on a computer-readable recording medium (for example, recording medium 23). As the recording medium, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like can be used. Magnetic disks include FD and HDD. Optical discs include CD, CD-R (Recordable) / RW (Rewritable), DVD, and DVD-R / RW. The program may be recorded and distributed on a portable recording medium. In that case, the program may be copied (installed) from a portable recording medium to another recording medium such as an HDD (for example, the HDD 103) and executed.

DESCRIPTION OF SYMBOLS 10 Information processing apparatus 11 Storage part 12 Operation part 13, 13a Evaluation function 14 Evaluated data 15 Evaluation value

Claims (7)

A storage unit for storing information of an evaluation function used for optimization processing;
Generate evaluation data, advance the optimization process on the evaluation data using the evaluation function, and calculate an evaluation value according to the state of the evaluation data after the optimization process And an arithmetic unit that updates the evaluation function based on an evaluation value according to a state of the evaluated data and the evaluated data after the optimization process is advanced ,
An information processing apparatus. The arithmetic unit, when updating the evaluation function, when a plurality of evaluated data can be used, an evaluation value according to a state of the evaluated data and a current evaluation after the optimization process is advanced Based on the difference with the value, select the evaluated data to be used for calculation from the plurality of evaluated data.
The information processing apparatus according to claim 1. The arithmetic unit advances the optimization process for other evaluated data using the updated evaluation function, and performs another process according to the state of the other evaluated data after the optimization process is advanced. Further updating the updated evaluation function by calculating an evaluation value,
The information processing apparatus according to claim 1 or 2. Information on the first evaluation function is stored in the storage unit,
When the calculation unit calculates a second evaluation function based on an evaluation value corresponding to a state of the evaluation target data and the evaluation target data after the optimization process is performed, the calculation unit uses the first evaluation function. Determining whether to update the first evaluation function to the second evaluation function according to a comparison between the evaluation result and the evaluation result by the second evaluation function;
The information processing apparatus according to any one of claims 1 to 3. The optimization process includes a multi-stage process,
The evaluation value according to the state of the evaluated data after the optimization process is advanced is up to the nth stage (n is an integer of 1 or more) in the middle of the optimization process with respect to the evaluated data. Calculated by executing the process,
The information processing apparatus according to any one of claims 1 to 4. An evaluation function learning method executed by a computer,
Generate evaluated data,
Advance the optimization process on the evaluated data using an evaluation function, calculate the evaluation value according to the state of the evaluated data after the optimization process,
Updating the evaluation function based on an evaluation value corresponding to a state of the evaluated data and the evaluated data after the optimization process is advanced ;
Evaluation function learning method. On the computer,
Generate evaluated data,
Advance the optimization process on the evaluated data using an evaluation function, calculate the evaluation value according to the state of the evaluated data after the optimization process,
Updating the evaluation function based on an evaluation value corresponding to a state of the evaluated data and the evaluated data after the optimization process is advanced ;
A program that executes processing.

Images