Optimization device and optimization method
By changing the values of four state variables in the Ising model to satisfy the two-way one-hot constraints, the problems of long calculation time and large search space in the prior art are solved, and a more efficient optimization process is achieved.
Patent Information
- Application Number
- CN202110275105.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-03-26
- Filing Date
- 2021-03-15
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-03-15
AI Technical Summary
When the existing Ising device deals with optimization problems with bidirectional one-heat constraints, the calculation time is long, and the conventional optimization device is easily converted to a state that does not meet the constraints during searching, which increases the search space and calculation time.
The two-way one-heat constraint is satisfied by changing the values of four state variables in the Ising model, the energy change of each state variable is calculated, and the state variables that are allowed to change are selected to shorten the calculation time and reduce the state transitions that do not satisfy the constraints. .
It effectively shortens the calculation time of optimization problems with two-way single-heat constraints, reduces the state transition that does not meet the constraints, and improves the efficiency of the optimization process.
Smart Images

Figure CN113449901B_ABST
Abstract
Description
Technical Field
[0001] This embodiment relates to an optimization device and an optimization method. Background Art
[0002] As a device for calculating large-scale discrete optimization problems that Neumann computers are not good at, there is an Ising device (also called a Boltzmann machine) that uses an Ising type evaluation function (also called an energy function, etc.).
[0003] In the calculation performed by the Ising device, the problem to be calculated is replaced by the Ising model, which is a model representing the spin behavior of a magnetic material. Then, a Markov chain Monte Carlo method such as a pseudo-fuzzy method or a replica exchange method (also called an exchange Monte Carlo method) is used to search for a state that minimizes the value of the Ising model-type evaluation function (corresponding to the energy of the Ising model).
[0004] Conventionally, there is an optimization device that searches for a state that minimizes energy by executing a Markov chain Monte Carlo method using a digital circuit. The conventional optimization device calculates the energy change amount (hereinafter may also be referred to as energy change) assuming that only one state variable value is changed at a time, and determines whether to allow a bit change based on a value obtained by adding a noise value corresponding to the temperature to the energy change. Changes in the state variable value that increase energy are also allowed with a predetermined probability, and the probability decreases as the temperature decreases.
[0005] By the way, some optimization problems have a constraint (one-hot constraint) in which the number of state variables having a value of 1 among a plurality of state variables included in the evaluation function is only one. There are some one-hot constraints in which when N 2 When the state variables are arranged in a matrix of N rows and N columns, the sum of the state variable values included in each row and each column becomes 1. In the following, this one-hot constraint is called a two-way one-hot constraint. For example, when x1 to x9 are arranged in a matrix of 3 rows and 3 columns, the constraints of x1+x2+x3=1, x4+x5+x6=1, x7+x8+x9=1, x1+x4+x7=1, x2+x5+x8=1, x3+x6+x9=1 are two-way one-hot constraints. The traveling salesman problem, vehicle allocation planning problem, quadratic allocation problem, etc. have two-way one-hot constraints. Summary of the invention
[0006] [Technical issues]
[0007] In a conventional optimization device, the number of state variables changed at one time is 1. For example, a conventional optimization device searches for a ground state with minimum energy while repeating a state transition with a Hamming distance of 1. Therefore, in a conventional optimization device, a transition to a state that does not satisfy the bidirectional one-hot constraint occurs, and the number of states in which the transition may occur (search space) is greater than the number of states that satisfy the bidirectional one-hot constraint. In addition, due to the energy barrier generated by the bidirectional one-hot constraint, state transitions require a certain amount of time. Based on the above content, a conventional optimization device has the following problem: it takes a certain amount of time to search for an optimal solution to an optimization problem with a bidirectional one-hot constraint.
[0008] In one aspect of the implementation, a solution is provided for reducing the computation time of an optimization problem with a bidirectional one-hot constraint.
[0009] [Solution to the problem]
[0010] According to one aspect of the embodiment, an optimization method implemented by an optimization device is provided, the method comprising: for each of a plurality of first state variables having a value of 0 before the change, calculating a change in the energy of an Ising model to satisfy the following constraint: when N (N is an integer of 2 or greater) square state variables included in an evaluation function representing the energy of the Ising model are arranged in N rows and N columns, the sum of the values of the state variables included in each row is 1 and the sum of the values of the state variables included in each column is 1, wherein the change is caused when the values of the four state variables are changed; selecting four state variables that are allowed to change based on the calculated change in each of the plurality of first state variables, and outputting identification information for identifying the selected four state variables; and updating the values of the four state variables that are allowed to change based on the identification information.
[0011] [Beneficial Effects of the Invention]
[0012] In one aspect of the implementation, the computation time of an optimization problem with a bidirectional one-hot constraint can be shortened. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 is a diagram showing an example of an optimization device according to the first embodiment;
[0014] Figure 2 is a diagram showing an example of an optimization device according to the second embodiment;
[0015] Figure 3 is a diagram showing an example of a propagation unit;
[0016] Figure 4 is a diagram showing an example of a propagation control unit;
[0017] Figure 5 is a diagram showing an example of state transitions satisfying a bidirectional one-hot constraint for twenty-five state variables;
[0018] Figure 6 It shows that what has happened Figure 5 A diagram showing an example of changes in the local fields and weighting coefficients used by each ΔE calculation circuit in the case of a state transition;
[0019] Figure 7 It shows h i and h j Figure 1 shows an example of the spread of ;
[0020] Figure 8 It shows h l Figure 1 shows an example of the spread of ;
[0021] Fig. 9 It shows h k Figure 1 shows an example of the spread of ;
[0022] Fig.10 It shows W il Figure 1 shows an example of the spread of ;
[0023] Fig.11 It shows W jk Figure 1 shows an example of the spread of ;
[0024] Fig.12 It shows the readout W il and storing it in a diagram of an example of a storage unit of a ΔE calculation circuit;
[0025] Fig.13 It shows the readout W jk and a diagram of an example of a storage unit stored in a ΔE calculation circuit;
[0026] Fig.14 It shows W jk Figure 1 shows an example of an address specification;
[0027] Fig.15 is a diagram showing an example of a storage unit storing 25×25 weighting coefficients;
[0028] Fig.16 is a diagram for describing a flow of an example of an optimization device according to a second embodiment;
[0029] Fig.17 is a diagram showing simulation results indicating the calculation shortening effect when the optimization device according to the second embodiment is used;
[0030] Fig.18 is a diagram showing an example of an optimization device according to a third embodiment;
[0031] Fig.19 is a diagram showing an example of a propagation control unit that propagates a local field and a weighting coefficient;
[0032] Fig. 20 is a diagram showing an example of energy change calculated in each cycle in the case of N=5;
[0033] Fig.21 is a diagram showing an operation example of the propagation control unit in cycle 1;
[0034] Fig. 22 is a diagram showing an operation example of the propagation control unit in cycle 2;
[0035] Fig.23 is a diagram showing an operation example of the propagation control unit in cycle 3;
[0036] Fig.24 is a diagram showing an example of the operation of the propagation control unit in cycle 4;
[0037] Fig.25 is a diagram showing an example of a propagation control unit in which propagation energy varies; and
[0038] Fig.26 is a diagram showing a modification of a ΔE calculation circuit. DETAILED DESCRIPTION
[0039] In a conventional optimization device, the number of state variables changed at one time is 1. For example, a conventional optimization device searches for a ground state with minimum energy while repeating a state transition with a Hamming distance = 1. Therefore, in a conventional optimization device, a transition to a state that does not satisfy the bidirectional one-hot constraint occurs, and the number of states in which the transition may occur (search space) is greater than the number of states that satisfy the bidirectional one-hot constraint. In addition, due to the energy barrier generated by the bidirectional one-hot constraint, state transitions require a certain amount of time. Based on the above content, a conventional optimization device has the following problem: it takes a certain amount of time to search for an optimal solution to an optimization problem with a bidirectional one-hot constraint.
[0040] In one aspect of the implementation, a solution is provided for reducing the computation time of an optimization problem with a bidirectional one-hot constraint.
[0041] Hereinafter, embodiments will be described with reference to the accompanying drawings.
[0042] (First Embodiment)
[0043] Figure 1 It is a diagram showing an example of the optimization device of the first embodiment.
[0044] The optimization apparatus 10 shortens the computation time of the optimization problem with the bidirectional one-hot constraint by excluding the search for states other than the states satisfying the bidirectional one-hot constraint.
[0045] The optimization problem is modeled in the Ising model. The energy of the modeled Ising model is defined by an evaluation function (E(x)) as shown in the following equation (1), for example.
[0046]
[0047] The first term on the right-hand side is obtained by integrating the product of the values of the two state variables and the weighting coefficients for all combinations of two state variables that can be selected from all the state variables included in the Ising model without omissions and duplications. i is a state variable with index = i. j is a state variable with index = j, W ij is a weighting coefficient indicating the weight (e.g., binding strength) between state variables with index = i and j. Note that W ii = 0. In addition, in many cases, W ij =W ji (For example, a coefficient matrix based on weighting coefficients is typically a symmetric matrix).
[0048] The second term on the right hand side is the sum of the products of the bias coefficients for each of all state variables and the state variable values. i indicates the bias coefficient of the state variable with index = i, and c is a constant.
[0049] For example, a spin value of -1 in the Ising model corresponds to a value of 0 in the state variable. A spin value of +1 in the Ising model corresponds to a value of 1 in the state variable. For this reason, the state variable can also be referred to as a bit that takes a value of 0 or 1.
[0050] The combination of values of the state variables that minimizes the value of equation (1) is the solution (optimal solution) to the problem.
[0051] Note that in equation (1), when x i The value of changes and becomes 1-x i When x i The increment can be expressed as Δx i =(1-x i )-x i =1-2x i The energy change associated with a change in this value (ΔE i ) is expressed by the following equation (2).
[0052]
[0053] In equation (2), when x i When Δx changes from 1 to 0, i becomes -1, and when x i When Δx changes from 0 to 1, i becomes 1. Note that h i is called the local field, and ΔE i is by adding h i Multiply by Δx i In other words, h i It means that due to x i The magnitude of the energy change in the Ising model caused by a change of 1.
[0054] When x j When h changes from 0 to 1 i The change in Δh i (j) =+W ij , and when x j When h changes from 1 to 0 i The change in Δh i (j) =-W ij . Similarly, when x i Change caused by x j h j The change in can be expressed as Δh j (i) =Δx i W ij .
[0055] Meanwhile, in the case of excluding the search for states other than the state satisfying the bidirectional one-hot constraint, the values of four state variables are changed in one state transition.
[0056] In the following, n=N included in the evaluation function 2 The state variables (N is an integer equal to or greater than 2) are subject to a bidirectional one-hot constraint.
[0057] By changing the values of four state variables in one state transition to satisfy the following constraint, the search for states other than the state satisfying the bidirectional one-hot constraint can be excluded: when the state variables are arranged in N rows and N columns, the sum of the values of the state variables included in each row and each column becomes 1.
[0058] Note that in the case where one of the state variables having a value of 0 is selected as a candidate to be updated in a state satisfying the bidirectional one-hot constraint, the other three candidate state variables to be updated are determined.
[0059] Figure 1shows that by changing four state variables (x i 、x j 、x k 、x l ) is an example of a state transition with a value of 0. j For the candidate case to be updated, as included in the x j x with value 1 in the state variable in the same row and column i and x l is the candidate to be updated. In addition, included in the i Same columns and x l The state variables in the same row have a value of 0 in x k A candidate to be updated.
[0060] ΔE j It can be expressed by the following equation (3), in which the energy change of the Ising model that occurs when the values of these four state variables are changed is ΔE j .
[0061] ΔE j =(h i +h l )-(h j +h k )-(W il +W jk ) (3)
[0062] In addition, since x i 、x j 、x k and x l The change in the local field caused by the change (Δh m (m=1, 2, ..., n)) can be expressed by the following equation (4).
[0063] Δh m =W jm +W km -(W im +W lm ) (4)
[0064] The optimization device 10 for changing the values of four state variables in one state transition includes, for example, Figure 1 The storage unit 11, propagation unit 12, calculation unit 13, conversion selection unit 14 and update unit 15, and address decoder 16 are shown. Figure 1 In FIG. 4 , a circuit for updating the local field according to equation (4), a control unit for controlling the size of a temperature parameter (T) to be described below, and the like are not shown.
[0065] The storage unit 11 stores a plurality of weighting coefficients included in the evaluation function of equation (1). The storage unit 11 is implemented using, for example, an electronic circuit such as a static random access memory (SRAM), a synchronous dynamic random access memory (SDRAM), or a high bandwidth memory (HBM).
[0066] The propagation unit 12 propagates N to each of the plurality of ΔE calculation circuits based on identification information to be described below output by the conversion selection unit 14. 2 Any of N local fields, 2 The local fields are to be used by the plurality of ΔE calculation circuits included in the calculation unit 13 to calculate the energy variation. Furthermore, the propagation unit 12 propagates weighting coefficients to be used by the plurality of ΔE calculation circuits for the energy variation to the corresponding plurality of ΔE calculation circuits based on the identification information.
[0067] The propagation unit 12 is implemented using electronic circuits such as one or more buses, a plurality of switches, and a control circuit for controlling the switches. An example of the propagation unit 12 will be described below.
[0068] For each of the multiple state variables having a value of 0 before the change, the calculation unit 13 calculates the energy change of the Ising model that occurs when the values of the four state variables are changed to satisfy the bidirectional one-heat constraint. The calculation unit 13 includes multiple ΔE calculation circuits (for example, ΔE calculation circuits 13ai and 13aj). For example, corresponding to N 2 state variables, set N 2 For example, the ΔE calculation circuit 13aj is at x j When ΔE is 0, calculate ΔE as shown in equation (3) j .
[0069] Furthermore, each of the plurality of ΔE calculation circuits may include a storage unit (eg, storage unit 13bi or 13bj) that reads in advance from the storage unit 11 a signal to be used for calculating ΔE as shown in equation (3). j The weighting factor (W il or W jk ) and stores the read weighting coefficients.
[0070] Note that in order to suppress the occurrence of meaningless state transitions, the ΔE calculation circuit corresponding to the state variable having a value of 1 outputs a predetermined positive value as the energy change. The predetermined positive value is, for example, the maximum positive value that the optimization device 10 can generate. For example, in the case where the optimization device 10 can generate a 26-bit value, when represented by a two's complement, the maximum positive value is 01...1 (the number of 1s is 25). Each ΔE calculation circuit reads the value of the state variable corresponding thereto from the storage unit 15a, and determines whether to calculate the energy change as shown in equation (3) or whether to output the above positive value.
[0071] The calculation unit 13 is realized using electronic circuits such as an adding circuit, a multiplying circuit (a circuit for multiplying a code of +1 or -1), and a memory (for example, a register or an SRAM).
[0072] The conversion selection unit 14 selects four state variables that are allowed to be changed based on the energy change calculated by the calculation unit 13, and outputs a signal for identifying the selected four state variables ( Figure 1 Identification information (index) of p, q, r and s in the example.
[0073] The transition selection unit 14 probabilistically accepts changes in the four state variables in a manner that gives priority to changes in the four state variables that cause negative energy changes when the values change among the plurality of energy changes output by the calculation unit 13. Changes in the four state variables that cause positive energy changes are also probabilistically permitted.
[0074] For example, the conversion selection unit 14 compares a noise value generated based on a temperature parameter (T) and a uniform random number input from a control unit (not shown) with each of a plurality of energy changes output by the calculation unit 13. In the case of executing the pseudo-fuzzy method, for example, each time the state update process of updating the Ising model is repeated by the control unit a predetermined number of times, T is controlled to have a smaller value, and the absolute value of the noise value also becomes smaller. The conversion selection unit 14 selects an energy change that is smaller than the noise value, and acquires an index of the energy change. For example, the conversion selection unit 14 selects ΔE j In the case of , the index = j is acquired. For example, the index is stored in each ΔE calculation circuit, and the conversion selection unit 14 may acquire the stored index, or the conversion selection unit 14 may generate the index. Note that in the case where there are a plurality of energy variations smaller than the noise value, the conversion selection unit 14 selects one of the energy variations, for example, according to a predetermined rule or randomly.
[0075] In the case of acquiring one index, the conversion selection unit 14 generates the other three indexes. For example, in the case of acquiring index=j, the conversion selection unit 14 generates the other three indexes as included in x. jx with value 1 in the state variable in the same row and same column i and x l The index = i and l, and the index included in x i Same columns and x i x in the same row k The index of = k. Note that k can be calculated by k = i + lj.
[0076] For example, the conversion selection unit 14 may generate the other three indexes based on the value of the state variable (identification information indicating which row or column the state variable belongs to is assigned to the state variable) stored in the storage unit 15a to be described below. In addition, the conversion selection unit 14 may store a hot bit management table that manages the index of the state variable (hot bit) having a value of 1 among the state variables in each row and each column, and may generate two indexes corresponding to the above i and l based on the table. In this case, the conversion selection unit 14 generates the index corresponding to the above k by the above equation (k=i+lj).
[0077] The conversion selection unit 14 is implemented using, for example, a random number generator, a circuit generating a noise value based on a temperature parameter (T), a comparator, a selector, a storage unit (such as a register or SRAM) storing a thermal bit management table, an adder, and the like.
[0078] The updating unit 15 includes storing x1 to x n The conversion selection unit 14 outputs the information to the update unit 15.
[0079] The updating unit 15 is implemented using, for example, an electronic circuit such as a circuit that inverts the value of a state variable specified by an index from 0 to 1 or from 1 to 0, a register, or an SRAM.
[0080] The address decoder 16 specifies the address of the storage unit 11 which stores the weighting coefficients to be used for calculating ΔE in the calculation unit 13 based on index=p, q, r, s. The address decoder 16 can be implemented using electronic circuits such as various logic circuits.
[0081] Note that the above elements may be mounted on a single semiconductor integrated circuit, or may be partially realized by one bit. For example, in the case where the scale of the problem is large, a storage unit 11 storing a large number of weighting coefficients may be provided outside the chip.
[0082] Hereinafter, an operation example of the optimization device 10 will be described.
[0083] First, an initial setting is performed. In the initial setting, for example, under the control of a control unit (not shown), x1 to x2 in N rows and N columns are nThe value of is set so that the sum of the values of the state variables included in each row and each column becomes 1. In addition, as an initial setting, execution will be performed according to x1 to x n The weighting coefficients of the initial values are stored in the storage units 13bi and 13bj, and the weighting coefficients are stored in the storage units 13bi and 13bj according to the weighting coefficients of the initial values. n The initial value sets the initial value of the local field, sets the temperature parameter, and sets the number of repetitions of the status update process.
[0084] Then, each ΔE calculation circuit calculates the energy change or outputs a predetermined positive value as described above, and the transition selection unit 14 selects four state variables allowed to change and outputs indexes=p, q, r, and s for identifying the selected four state variables.
[0085] The updating unit 15 receives indexes = p, q, r and s, and changes the values of the state variables specified by these indexes. In addition, the address decoder 16 specifies the address of the storage unit 11 storing the weighting coefficients to be used for the next calculation of energy change and the weighting coefficients for updating the local field based on the indexes, so that these weighting coefficients can be read out.
[0086] The weighting coefficients to be used for calculating the energy variation among the read weighting coefficients are stored in any of the storage units of the plurality of ΔE calculation circuits via the propagation unit 12. Furthermore, the local field is updated using the read weighting coefficients for updating the local field, and the updated local field is propagated to any of the plurality of ΔE calculation circuits through the propagation unit 12. Then, each ΔE calculation circuit calculates the energy variation or outputs the above predetermined positive value.
[0087] For example, in x j is 0, and x j x in the same row i is 1, and x j x in the same column l is 1, and x i Same column as x l x in the same row k When W is 0, il and W jk is propagated and stored in the storage unit 13bj of the ΔE calculation circuit 13aj. i 、h j 、h k and h l is propagated to the ΔE calculation circuit 13aj. Then, the ΔE calculation circuit 13aj calculates ΔE as shown in equation (3): j . In the case of another state variable (x k Similar processing is performed in the corresponding ΔE calculation circuit (etc.).
[0088] At the same time, with x having a value of 1 i The corresponding ΔE calculation circuit 13ai outputs a predetermined positive value as the energy change. Note that, unlike another state variable (x l The corresponding ΔE calculation circuit similarly outputs a predetermined positive value.
[0089] In the following, similar processing is repeated. For example, in the case of executing the pseudo-fuzzy method, each time the state update process is completed for a predetermined number of repetitions, the value of the temperature parameter becomes smaller according to a predetermined temperature change schedule. Then, for example, the state (x1 to x2) when the value of the temperature parameter reaches the minimum value is set. n ) as the solution output.
[0090] The optimization device 10 according to the first embodiment determines which four state variables are allowed to change based on the energy change when the four state variables change together to satisfy the bidirectional one-hot constraint. Then, the values of the determined four state variables are updated. Therefore, the state transition that does not satisfy the bidirectional one-hot constraint is suppressed, and the search space can be reduced. Therefore, the calculation time of the optimization problem with the bidirectional one-hot constraint can be shortened.
[0091] Furthermore, since constraint terms that increase energy when the bidirectional one-hot constraint is not satisfied can be reduced from the evaluation function, the number of bits of weighting coefficients used to represent such constraint terms can be reduced, and hardware used to store the weighting coefficients can be reduced.
[0092] (Second Embodiment)
[0093] Figure 2 It is a diagram showing an example of the optimization device of the second embodiment.
[0094] The optimization device 20 of the second embodiment includes a storage unit 21, a local field update unit 22, a propagation unit 23, ΔE calculation circuits 24a1 to 24an, a conversion selection unit 25, an update unit 26, an address decoder 27 and a control unit 28.
[0095] The storage unit 21, the propagation unit 23, the ΔE calculation circuits 24a1 to 24an, the conversion selection unit 25, the update unit 26 and the address decoder 27 have Figure 1 Note that in the following example, in addition to the indexes = p, q, r, and s, the conversion selection unit 25 also outputs a function for identifying x having a value of 1. p The index of the row to which it belongs = g p and is used to identify x with a value of 1 s Index of the column to which it belongs = g s .
[0096] The local field updating unit 22 updates the local field according to equation (4) using the weighting coefficients read from the storage unit 21. Using, for example, an electronic circuit such as a storage unit 21 storing the weighting coefficients corresponding to x1 to x2, the local field updating unit 22 updates the local field according to equation (4). n Corresponding h1 to h n resistors, and the stored h1 to h n Add four weighting coefficients or from the stored h1 to h n The local field updating unit 22 is implemented by a circuit that subtracts four weighting coefficients respectively.
[0097] The control unit 28 performs the initial setting processing to be described below for the optimization device 20. In addition, in the case of executing the pseudo-fuzzy method, for example, each time the state update processing is repeated a predetermined number of times, the control unit 28 makes the value of the temperature parameter smaller according to the temperature change schedule specified by the control device 30.
[0098] Furthermore, for example, the control unit 28 acquires the state (x1 to x2) when the value of the temperature parameter reaches the minimum value from the storage unit 26a of the updating unit 26. n ), and transmits the acquired state as a solution to the control device 30. Note that in the case where the storage unit 26a of the update unit 26 stores the minimum energy or the state at the minimum energy, the control unit 28 can obtain its information and transmit the information to the control device 30 after repeating the state update process a predetermined number of times.
[0099] The control unit 28 may be implemented by an electronic circuit such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc. Note that the control unit 28 may be a processor such as a central processing unit (CPU) or a graphics processing unit (GPU). In this case, the processor performs the above-mentioned processing by executing a program stored in a memory (not shown).
[0100] (Example of propagation unit 23)
[0101] Figure 3 is a diagram showing an example of a propagation unit. Note that Figure 3 An example of a local field update unit 22 is also shown. Figure 1 The propagation unit 12 shown in FIG. 1 can also be implemented by a configuration similar to the example of the propagation unit 23 to be described below.
[0102] The propagation unit 23 includes a switch control circuit 23 a , switches 21 b 1 to 21 bn , and a propagation control unit 23 c .
[0103] The switch control circuit 23a is based on the index = p, q, r, s, g p and g sTo control the on / off of the switches 21b1 to 21bn. The switch control circuit 23a may include a memory for storing the thermal level management table as described above. In this case, the switch control circuit 23a generates control signals for the switches 21b1 to 21bn based on the thermal level management table. The thermal level management table is based on indexes = p, q, r, s, g p and g s Next, a control example of the switches 21b1 to 21bn will be described.
[0104] When the switches 21 b 1 to 21 bn are turned on, the weighting coefficients read from the storage unit 21 are transmitted to the propagation control unit 23 c .
[0105] The propagation control unit 23c is based on the index = p, q, r, s, g p and g s , h1 to h2 are outputted by h updating circuits 22a1 to 22an included in the local field updating unit 22. n Any of the above is propagated to each of the ΔE calculation circuits 24a1 to 24an. In addition, the propagation control unit 23c is based on the index = p, q, r, s, g p and g s , propagates any of the weighting coefficients supplied via the switches 21b1 to 21bn to each of the ΔE calculation circuits 24a1 to 24an.
[0106] Figure 4 is a diagram showing an example of a propagation control unit.
[0107] The propagation control unit 23 c includes bus units 40 a 1 , 40 a 2 , . . . , 40 am and switch control circuits 41 and 42 .
[0108] The bus unit 40a1 includes switches 40b1, 40b2, ..., 40bn, 40c1, 40c2, ..., and 40cn and a bus 40d. The switches 40b1 to 40bn are connected to h1 to h2. n The switches 40b1 to 40bn are provided correspondingly, and when any one of the switches 40b1 to 40bn is in the on state, the local field corresponding to the switch is transmitted to the bus 40d. The switches 40c1 to 40cn are provided correspondingly to the ΔE calculation circuits 24a1, 24a2, ..., and 24an, and the local field is propagated to the ΔE calculation circuit corresponding to the switch in the on state among the switches 40c1 to 40cn via the bus 40d. The bus units 40a2 to 40am have a configuration similar to that of the bus unit 40a1.
[0109] The switch control circuit 41 is based on the index = p, q, r, s, g p and g sTo control the on / off of switches 40b1 to 40bn.
[0110] The switch control circuit 42 is based on the index = p, q, r, s, g p and g s To control the on / off of switches 40c1 to 40cn.
[0111] Note that the propagation control unit 23c may include a memory for storing the thermal level management table as described above. In this case, the switch control circuits 41 and 42 generate control signals for each of the switches based on the thermal level management table. The thermal level management table is based on index = p, q, r, s, g p and g s Note that the thermal level management table may be the same as the thermal level management table used by the switch control circuit 23a or the like.
[0112] In the above example, the number of bus units 40a1 to 40am can be appropriately selected according to the degree of priority given to the degree of parallelism in calculation.
[0113] Notice, Figure 4 An example is shown in which the local fields are propagated to the ΔE calculation circuits 24a1 to 24an. The weighting coefficients can also be propagated to the ΔE calculation circuits 24a1 to 24an by a similar configuration.
[0114] (Example of propagation of local fields and weighting coefficients when n=25)
[0115] Hereinafter, propagation of the local fields and weighting coefficients to the ΔE calculation circuits 24 a 1 to 24 an will be described taking a case where n=25 (the number of state variables is twenty-five) as an example.
[0116] Figure 5 is a diagram showing an example of state transitions satisfying a bidirectional one-hot constraint for twenty-five state variables. Figure 5 In order to simplify the explanation, x1 to x 25 Indicated by index=1 to 25, g1c to g5c are indexes for identifying columns, and g1r to g5r are indexes for identifying rows. In addition, shaded indexes indicate indexes of state variables having a value of 1, and values of other state variables are 0.
[0117] exist Figure 5 In the example, x1, x8, x 15 、x 17 and x 24 The value of is 1. In this example, when x4 is changed from 0 to 1, x belonging to the same column as x4 (the column with index=g4c) and having a value of 1 24The value of is changed to 0, and the value of x1, which belongs to the same row as x4 (row with index = g1r) and has a value of 1, is changed to 0. In addition, x1, which belongs to the same column as x1 (column with index = g1c) and the same row as x24 (row with index = g5r), is changed to 0. 21 The value of is changed to 1. Thus, the state transition that satisfies the bidirectional unique-hot constraint is achieved.
[0118] In the occurrence of Figure 5 In the case of the state transition shown, the energy change before and after the transition is calculated by the ΔE calculation circuits 24a1 to 24an using the local fields and the weighting coefficients.
[0119] Figure 6 It shows that what has happened Figure 5 A diagram of an example of changes in the local fields and weighting coefficients used by each ΔE calculation circuit in the case of a state transition. Figure 6 Shows Figure 5 The h before and after the state transition are shown in i 、h j 、h k 、h l , W il and W jk , which are used when each of the twenty-five ΔE calculation circuits 24a1 to 24an calculates equation (3). Note that in order to simplify the description, the local fields and weighting coefficients are shown by index. For example, h1 is shown as 1, and W 11 Shown as (1,1).
[0120] exist Figure 6 In the example of , before the state transition, the h propagated to the ΔE calculation circuit corresponding to the hot bits such as x1 and x8 i 、h j 、h k and h l is a local field with the same index as the hot bit. For example, h is provided to the ΔE calculation circuit 24a1 corresponding to x1 as the hot bit. i 、h j 、h k and h l are both h1. Therefore, W il and W jk All W 11 As described above, the ΔE calculation circuit corresponding to such a thermal level outputs a predetermined positive value.
[0121] After the state transition, in the five ΔE calculation circuits corresponding to x1 to x5 belonging to the row with index = g1r, x4 belonging to the same row changes from 0 to 1, and therefore, h4 is used as h1 instead of h1 before the state transition.i . In the case of x6 to x belonging to the row with index = g2r 10 In the corresponding five ΔE calculation circuits, since the hot bit belonging to the row is not changed, h8 is used as h as before the state transition. i In the ΔE calculation circuit corresponding to the state variables belonging to the rows with index = g3r and g4r, since the hot bits belonging to those rows are not changed, the local field with the same index as before the state transition is used as h i . In the case of x belonging to the row with index = g5r 21 to x 25 In the corresponding five ΔE calculation circuits, x belonging to the same row 21 changes from 0 to 1, thus replacing the h before the state transition 24 ,h 21 Used as h i .
[0122] Before and after the state transition, h j No changes.
[0123] About h k , in the ΔE calculation circuit corresponding to the state variable belonging to the row without hot bit change, it is possible that h changes before and after the state transition. For example, h used by the ΔE calculation circuit corresponding to x6 k Change from h3 to h 23 In addition, h used by the ΔE calculation circuit corresponding to x9 k From h 23 Change to h3.
[0124] About h l , in the ΔE calculation circuit corresponding to the state variable belonging to each row, there may be changes before and after the state transition. Note that with respect to h l , h used in the ΔE calculation circuit corresponding to the state variable belonging to each row l That is, h used in the ΔE calculation circuit corresponding to the state variable belonging to each row l Before the state transition, it is h1, h 17 、h8、h 24 and h 15 , and after the state transition is h 21 、h 17 , h8, h4 and h 15 .
[0125] As for the weighting coefficients to be used, as described above, some weighting coefficients change and some weighting coefficients do not change depending on the change of the local fields to be used before and after the state transition.
[0126] Hereinafter, an example of propagation of the local field and the weighting coefficient to each ΔE calculation circuit after state transition will be described.
[0127] Figure 7 It shows h i and h j Figure 2 shows an example of the propagation of .
[0128] Figure 7 The h1 to h2 outputted by the h updating circuits 22a1 to 22an are shown. 25 As h i and h j The state is propagated to the ΔE calculation circuits 24a1 to 24an.
[0129] h1 to h 25 Each of the h j is propagated to the ΔE calculation circuit corresponding to the state variable with the same index as h. Before and after the state transition, the propagated h j For example, h1 is propagated to the ΔE calculation circuit 24a1 corresponding to x1, and h n is propagated to x n Therefore, if the ΔE calculation circuits corresponding to the h update circuits 22a1 to 22an are directly connected, h can be propagated without using the propagation control unit 23c. j .
[0130] The propagation control unit 23c is used to propagate h as follows i .
[0131] h4 as h i is propagated to the five ΔE calculation circuits corresponding to x1 to x5 belonging to the row with index = g1r, and h8 is used as h i is propagated to x6 to x belonging to the row with index = g2r 10 The corresponding five ΔE calculation circuits. 15 As h i is propagated to x belonging to the row with index = g3r 11 to x 15 The corresponding five ΔE calculation circuits, and h 17 As h i Propagate to x belonging to the row with index = g4r 16 to x 20 The corresponding five ΔE calculation circuits. In addition, h 21 As h i Propagate to x belonging to the row with index = g5r 21 to x 25 The corresponding five ΔE calculation circuits.
[0132] In the above-mentioned propagation, Figure 4 The propagation control unit 23c shown in FIG. 1 can convert h into m in one cycle (for example, one clock cycle) when m=5. i For example, the bus unit 40a1 transmits h4 to the bus 40d, and propagates h4 from the bus 40d to the five ΔE calculation circuits corresponding to x1 to x5 belonging to the row with index = g1r. In parallel, the bus units 40a2 to 40am transmit h8, h9, and h10 to the five ΔE calculation circuits corresponding to x1 to x5 belonging to the row with index = g1r. 15 、h 17 and h 21 Propagate to the ΔE calculation circuits corresponding to the state variables belonging to the rows with index = g2r to g5r.
[0133] Figure 8 It shows h l Figure 2 shows an example of the propagation of .
[0134] like Figure 8 As shown above, as h l Spread h4, h8, h 15 、h 17 and h 21 Each of them is h l Propagates to any of the ΔE calculation circuits 24 a 1 to 24 an .
[0135] h4 is propagated to x4 and x9, x 14 、x 19 and x 24 The corresponding five ΔE calculation circuits. h8 is propagated to x8 and x3, x4, and x5 that belong to the same column as x8. 13 、x 18 and x 23 The corresponding five ΔE calculation circuits. 15 is propagated to x 15 and x 15 x5, x2, and x3 belong to the same column. 10 、x 20 and x 25 The corresponding five ΔE calculation circuits. 17 is propagated to x 17 and x 17 x2, x7, x that belong to the same column 12 and x 22 The corresponding five ΔE calculation circuits. 21 is propagated to x 21 and x 21 x1, x6, x that belong to the same column 11 and x16 The corresponding five ΔE calculation circuits.
[0136] It is also possible to use five bus units 40a1 to 40am to execute the above-described operations in parallel in one cycle. l of the spread.
[0137] Fig. 9 It shows h k Figure 2 shows an example of the propagation of .
[0138] As for h k , in such Figure 5 The state after the state transition shown is the same as the row and column to which x4 belongs and x 21 The local fields corresponding to the state variables in the corresponding rows and columns are taken as h k To spread.
[0139] Using, for example, five bus units 40a1 to 40am, h is executed in four cycles. k of the spread.
[0140] In cycle c1, each of h1 to h5 is transferred to one of the five buses in bus units 40a1 to 40am, and h1 to h5 are propagated to the bus with x 24 、x 19 , x9, x4 and x 14 The corresponding ΔE calculation circuits. In period c2, h1, h6, h 11 、h 16 and h 21 Each of them is transmitted to one of the five buses, and h1, h6, h 11 、h 16 and h 21 is transmitted to the 24 、x 23 、x 25 、x 22 and x 21 The corresponding ΔE calculation circuits. In period c3, h4, h9, h 14 、h 19 and h 24 Each of them is transmitted to any of the five buses, and h4, h9, h 14 、h 19 and h 24 is propagated to the ΔE calculation circuits corresponding to x4, x3, x5, x2, and x1, respectively. In cycle c4, h 21 、h 22 、h 23 、h 24 and h 25 Each of them is transmitted to any of the five buses, and h21 、h 22 、h 23 、h 24 and h 25 is propagated to x 21 、x 16 , x6, x1, and x 11 The corresponding ΔE calculation circuits respectively.
[0141] Note that as h k The number of local fields propagated is essentially fourteen: h1, h2, h3, h5, h6, h9, h 11 、h 14 、h 16 、h 19 、h 22 、h 23 、h 24 and h 25 This is because h4 and h 21 As h i is propagated to the hot positions x4 and x 21 The corresponding ΔE calculation circuit. Therefore, if the propagation of these redundancies is reduced, then h k Can ideally be spread over three cycles.
[0142] Fig.10 It shows W il Figure 2 shows an example of the propagation of .
[0143] like Figure 6 As shown, twenty-five W are used in the twenty-five ΔE calculation circuits 24a1 to 24an after the state transition. il , and twenty-five W il Nine of them are the same as before the state transition. In addition, when W il =W li When W is to be propagated il Therefore, for example, five bus units 40a1 to 40am can be used to perform W in two cycles. il spread.
[0144] In the first cycle, W 4,8 , W 4,15 , W 4,17 and W 4,21 is propagated to x3 and x9, x5 and x 14 , x2 and x 19 and x1 and x 24 Note that since the ΔE calculation circuit corresponding to x4 outputs a predetermined positive value, it is not necessary to set W 4,4 propagates to the ΔE calculation circuit. Therefore, Fig.10In the middle, W is not propagated 4,4 However, due to the addition to the propagation W 4,4 The processing overhead is small, so W can be propagated 4,4 For example, since the circuit configuration of the propagation control unit 23c is simplified, there may be a desired W 4,4 possibility of transmission.
[0145] In the second cycle, W 21,8 , W 21,15 and W 21,17 is propagated to x6 and x 23 、x 11 and x 25 and x 16 and x 22 The corresponding ΔE calculation circuits are shown in Figure 1. Note that due to the relationship between x 21 The corresponding ΔE calculation circuit outputs a predetermined positive value, so there is no need to change W 21,21 propagates to the ΔE calculation circuit. Therefore, Fig.10 In the middle, W is not propagated 21,21 . Note that W 21,21 Can be spread for similar reasons as above.
[0146] Fig.11 It shows W jk Figure 2 shows an example of the propagation of .
[0147] like Figure 6 As shown, twenty-five W are used in the twenty-five ΔE calculation circuits 24a1 to 24an after the state transition. jk , and twenty-five W jk Nine of them are the same as before the state transition. In addition, when W il =W li When W is to be propagated jk Therefore, for example, five bus units 40a1 to 40am can be used to perform W in two cycles. jk spread.
[0148] In the first cycle, W 3,9 , W 5,14 , W 2,19 and W 1,24 Propagate to x3 and x9, x5 and x 14 , x2 and x 19 and x1 and x 24 The corresponding ΔE calculation circuits respectively.
[0149] In the second cycle, W 23,6 , W 25,11 and W 22,16 Propagate to x6 and x 23、x 11 and x 25 and x 16 and x 22 The corresponding ΔE calculation circuits are shown in Figure 1. Note that due to similar reasons as above, Fig.11 W is not propagated 4,4 and W 21,21 .
[0150] (Example of reading weighting coefficients when n=25)
[0151] Next, an efficient method of reading the weighting coefficients from the storage unit 21 will be described.
[0152] In such Figure 5 After the state transition shown, read W 1,1 To W 1,25 , W 4,1 To W 4,25 , W 21,1 To W 21,25 , and W 24,1 To W 24,25 To update the local field of 25×25 weighting coefficients stored in the storage unit 21. 1,1 To W 1,25 , W 4,1 To W 4,25 , W 21,1 To W 21,25 , and W 24,1 To W 24,25 In, W 4,1 To W 4,25 , and W 21,1 To W 21,25 Contains the W to be used il .
[0153] Fig.12 It shows the readout W il And a diagram of an example of a storage unit stored in a ΔE calculation circuit.
[0154] exist Fig.12 In the example of 4,1 To W 4,25 To update the local field. Thus, for example, as W il , W 4,8 is provided to the h update circuit 22a8 to update h8, and W 4,15 is provided to the h update circuit 22a15 to update h 15 In addition, for example, as W il , W 4,17 is provided to the h update circuit 22a17 to update h 17 , and W 4,21is provided to the h update circuit 22a21 to update h 21 .
[0155] At this time, the switch control circuit 23a turns on x8 and x9 whose values after the state transition are 1, based on, for example, the hot bit management table. 15 、x 17 and x 21 The corresponding switches 21b8, 21b15, 21b17 and 21b21 are turned off. Although not shown, the other switches are turned off. 4,8 , W 4,15 , W 4,17 and W 4,21 is provided to the propagation control unit 23c. Note that the control of W may be omitted for similar reasons as above. 4,4 The reading is therefore Fig.12 In the example, W is not 4,4 Supplied to the propagation control unit 23c.
[0156] As in Fig.10 As in the first cycle of 4,8 , W 4,15 , W 4,17 and W 4,21 Each of is propagated to any of the ΔE calculation circuits 24 a 1 to 24 an .
[0157] For example, Fig.12 As shown, W 4,21 The W is propagated to the ΔE calculation circuit 24a1 and stored in the storage areas 24b1a and 24b1b of the storage unit 24b1. il The storage area 24b1a of W 4,17 The W is propagated to the ΔE calculation circuit 24a2 and stored in the storage areas 24b2a and 24b2b of the storage unit 24b2. il In addition, W 4,8 The W is propagated to the ΔE calculation circuit 24a3 and stored in the storage areas 24b3a and 24b3b of the storage unit 24b3. il in the storage area 24b3a.
[0158] Similarly, when reading W from storage unit 21 21,1 To W 21,25 In the case of updating the local field, switches 21b8, 21b15, 21b17 and 21b21 are turned on. 21,8 , W 21,15 , W 21,17 and W 21,21is provided to the propagation control unit 23c. Note that for similar reasons to the above, W may be omitted. 21,21 reading and dissemination.
[0159] Then, W is supplied to the propagation control unit 23c. 21,8 , W 21,15 and W 21,17 Each of the ΔE calculation circuits 24a1 to 24an is propagated to any of the ΔE calculation circuits 24a1 to 24an and stored in the storage unit for storing W il In the storage area, such as Fig.10 The same is true in the second cycle.
[0160] As described above, W can be read from the storage unit 21 at the same timing as the readout of the weighting coefficient for updating the local field. il , and there is no need to read W from storage unit 21 again il .
[0161] At the same time, in Figure 5 After the state transition shown, except W 1,24 In addition, W jk is not included in the weighting coefficients read for updating the local fields. 1,1 To W 1,25 , W 4,1 To W 4,25 , W 21,1 To W 21,25 and W 24,1 To W 24,25 During the read cycle, read the W 1,24 W jk .
[0162] Fig.13 It shows the readout W jk And a diagram of an example of a storage unit stored in a ΔE calculation circuit.
[0163] like Figure 6 As shown, twenty-five W are used in the twenty-five ΔE calculation circuits 24a1 to 24an after the state transition. jk , and twenty-five W jk Nine of them are the same as before the state transition. In addition, when W il =W li When W is to be read from storage unit 21 jk The number is nine. However, for the reasons given above, it is not necessary to read W 4,4 and W 21,21 , and in reading W 1,1 To W 1,25 When used to update local fields, W 1,24 is stored using W1,24 The ΔE calculation circuit is used for the weighting coefficient storage unit for storing W jk Therefore, the data to be read from the storage unit 21 as W jk The number of weighting coefficients is six: W 23,6 , W 3,9 , W 25,11 , W 5,14 , W 22,16 and W 2,19 By the configuration of the address decoder 27 and the storage unit 21 to be described below, these weighting coefficients can be read in one cycle.
[0164] The switch control circuit 23a controls the state variable (x) having a value of 0 after the state transition based on, for example, the thermal level management table. j ), detection such as Figure 1 The x shown k state variables in the relation, and detects the j With x k W between jk In the above example, k=6, 9, 11, 14, 16, and 19 are detected. Therefore, the switch control circuit 23a turns on the switches 21b6, 21b9, 21b11, 21b14, 21b16, and 21b19. The switches 21b6, 21b9, 21b11, 21b14, 21b16, and 21b19 are set corresponding to the h update circuits 22a6, 22a9, 22a11, 22a14, 22a16, and 22a19. Although not shown, the other switches are turned off.
[0165] Thus, the above six weighting coefficients are supplied to the propagation control unit 23c.
[0166] Propagation of six weighting coefficients is performed in two cycles using five bus units 40a1 to 40am, for example, of the propagation control unit 23c.
[0167] For example, Fig.13 As shown, W 2,19 is propagated to the ΔE calculation circuit 24a2 and stored in the memory cell 24b2 for storing W jk For example, W 3,9 is propagated to the ΔE calculation circuit 24a3 and stored in the memory cell 24b3 for storing W jk in the storage area 24b3b.
[0168] (Example of Storage Unit 21 and Address Specification Example)
[0169] In the following example, the weighting coefficients (W ij (i, j = 1 to 25)) is specified by a global row address, a local row address, a global column address, and a local column address. Every five rows of the weighting coefficients of the twenty-five rows are specified by each of the different global row addresses = 1 to 5, and the five rows are respectively specified by the local row addresses = 1 to 5. In addition, every five columns of the weighting coefficients of the twenty-five columns are specified by each of the different global column addresses = 1 to 5, and the five columns are respectively specified by the local column addresses = 1 to 5.
[0170] By specifying one of 1 to 5 as the global row address and the local row address, and specifying all of 1 to 5 as the global column address and the local column address, the weight coefficient to be used for updating the local field is read. At the same time, the above-mentioned W is read by specifying an address such as described below jk .
[0171] Fig.14 It shows W jk Figure 2 is an example of an address specification.
[0172] For example, W 2,19 Specified by global row address = 1, local row address = 2, global column address = 4, and local column address = 4. 3,9 Specified by global row address = 1, local row address = 3, global column address = 2, and local column address = 4. The other four W jk Also specified by four different addresses.
[0173] Fig.15 is a diagram showing an example of a storage unit storing 25×25 weighting coefficients.
[0174] Fig.15 The method for selecting a w-bit weighting coefficient (W 11 ). The signal grad1 is a signal that becomes 1 when the global row address is 1 and becomes 0 in other cases, and the signal lrad1 is a signal that becomes 1 when the local row address is 1 and becomes 0 in other cases. In addition, the signal gcad1 is a signal that becomes 1 when the global column address is 1 and becomes 0 in other cases, and the signal lcad1 is a signal that becomes 1 when the local column address is 1 and becomes 0 in other cases.
[0175] wW 11 Stored in the memory cells 21a1 to 21aw. The signals grad1, lrad1, and gcad1 are input to a 3-input AND (logical product) circuit 21b.
[0176] The output signal of the 3-input AND circuit 21b and the signal lcad1 are input to the 2-input AND circuit 21c, and when the output of the 2-input AND circuit 21c is 1, the memory cells 21a1 to 21aw are selected. Then, the read circuit 21e reads W via the bit lines 21d1 to 21dw. 11 .
[0177] Five units each including the memory units 21a1 to 21aw, the 2-input AND circuit 21c, and the bit lines 21d1 to 21dw as described above are connected to the signal line on which the output signal of the 3-input AND circuit 21b propagates. In addition, five units each including the above five units and the 3-input AND circuit 21b are connected to the signal line on which the signal lrad1 propagates. In addition, this configuration is repeatedly provided for five local row addresses, and the configuration provided for the five local row addresses is further repeatedly provided for five global row addresses.
[0178] The signal line for each of the above addresses can be independently set to 1 or 0. In addition, if the bidirectional single-hot constraint is satisfied, different memory cells connected to the same bit line will not be selected at the same time. Therefore, the above six W jk , and the above six W can be read in one cycle jk .
[0179] Figure 2 The address decoder 27 in is based on the index = p, q, r, s, g p and g s Generates an address for specifying the weight coefficients to be read. Generates addresses for specifying all weight coefficients for rows p, q, r, and s when updating local fields. jk When the address decoder 27 is based on the index = p, q, r, s, g p and g s Select multiple W to read jk , and generate the jk For example, in Figure 5 After the state transition shown, index = q, and r = 4 and 21. In this case, address decoder 27 generates the address for reading six: 23,6 , W 3,9 , W 25,11 , W 5,14 , W 22,16 and W 2,19 As W jk The address is as above.
[0180] The address decoder 27 may include a memory for storing a hot bit management table as described above. In that case, the address decoder 27 is based on the index = p, q, r, s, g p and g s Select multiple W to read jk The thermal level management table may be the same as the table used by the switch control circuit 23a or the like.
[0181] The global row address can be expressed as Gr(j)=int((j-1) / size)+1, and the local row address can be expressed as GREL(j)=mod(j-1, size)+1. In addition, the global column address can be expressed as Gr(k)=int((k-1) / size)+1, and the local column address can be expressed as GREL(k)=mod(k-1, size)+1.
[0182] In the case where each address is set to 1 to 5 as described above, the size is 5.
[0183] For example, when reading W 23,6 In the case of , Gr(23) = int((23-1) / 5)+1 = 5, GREL(23) = mod(23-1,5)+1 = 3, Gr(6)) = Int((6-1) / 5)+1 = 2, and GREL(6) = mod(6-1,5)+1 = 1. When reading W 3,9 In this case, Gr(3)=int((3-1) / 5)+1=1, GREL(3)=mod(3-1,5)+1=3, Gr(9)=Int((9-1) / 5)+1=2, and GREL(9)=mod(9-1,5)+1=4.
[0184] As mentioned above, in Fig.15 In the illustrated storage unit 21 , the signal line for each address can be independently set to 1 or 0, so that the address decoder 27 simultaneously sets various addresses as described above.
[0185] (Overall Operation Flow of the Optimization Device 20 of the Second Embodiment)
[0186] Fig.16 It is a diagram for describing a flow of an example of the optimization device according to the second embodiment.
[0187] Note that, hereinafter, description will be given using the pseudo-blur method as an example. However, the embodiment is not limited to this case, and a technique such as a replica exchange method may also be used.
[0188] First, the initial setting process is performed under the control of the control unit 28 (step S1). In the initial setting process, the control unit 28 receives the W received from the control device 30.11 To W nn is stored in the storage unit 21. In addition, the control unit 28 sets x1 to x2 in N rows and N columns in the storage unit 26a. n The value of x1 to x2 is such that the sum of the values of the state variables included in each row and each column is 1. In addition, for example, the control unit 28 will n The weighting coefficients corresponding to the initial values of x1 to x2 are stored in the storage units 24b1 to 24bn and are n The initial value of h1 to h n The initial value of h is stored in the resistors (not shown) of the h update circuits 22a1 to 22an. In addition, the control unit 28 sets the initial value of the temperature parameter (T) based on the temperature change schedule received from the control device 30 and the number of repetitions of the state update process of the conversion selection unit 25. In addition, for example, the control unit 28 can set the initial value of the temperature parameter (T) according to x1 to x2. n The above-mentioned heat level management table is generated based on the initial value of , and the heat level management table is stored in a memory included in the propagation unit 23 or the like.
[0189] After that, the ΔE calculation circuits 24a1 to 24an read the state variables corresponding to themselves from the storage unit 26a (step S2). Then, the propagation unit 23 converts h into i 、h j 、h k and h l The h updating circuits 22a1 to 22an are propagated to the ΔE calculation circuits 24a1 to 24an (step S3). First, the propagation unit 23 determines h based on, for example, the thermal level management table. i 、h j 、h k and h l Then, from the second time onwards, the propagation unit 23 uses the index = p, q, r, s, g p and g s Or use the heat management table updated with these indexes to determine the propagation route.
[0190] Next, the ΔE calculation circuits 24a1 to 24an calculate the energy change (step S4). Among the ΔE calculation circuits 24a1 to 24an, those corresponding to the state variable having a value of 0 calculate ΔE as shown in equation (3): j Among the ΔE calculation circuits 24a1 to 24an, those corresponding to the state variable having a value of 1 output a predetermined positive value as the energy change.
[0191] Then, the conversion selection unit 25 selects index=q based on the energy changes calculated by the ΔE calculation circuits 24 a 1 to 24 an (step S5 ).
[0192] For example, the conversion selection unit 25 compares the noise value generated based on the temperature parameter (T) and the uniform random number with each of the multiple energy changes, selects the energy change that is smaller than the noise value, and selects the index of the energy change as q. In the case where there are multiple energy changes that are smaller than the noise value, the conversion selection unit 25 selects one of the energy changes, for example, according to a predetermined rule or randomly. In the case where there is no energy change that is smaller than the noise value, the conversion selection unit 25 can promote the occurrence of state conversion by adding an offset value to the noise value, etc.
[0193] The conversion selection unit 25 also determines the indexes = p, q, r, s, g according to the selected index = q p and g s (Step S6).
[0194] Thereafter, the local field updating unit 22 updates the local field (step S7). In the process in step S7, based on the index = p, q, r, s, g p and g s , read the weighting coefficient to be used for updating the local field from the storage unit 21, and update the local field based on the weighting coefficient according to equation (4). Note that in the processing in step S7, the weighting coefficients are il The corresponding weighting coefficient is stored in any of the storage units 24b1 to 25bn.
[0195] In addition, the weighting coefficients stored in the storage units 24b1 to 24bn of the ΔE calculation circuits 24a1 to 24an are updated (step S8). In the process of step S8, based on the index = p, q, r, s, g p and g s , the weighting coefficients stored in the storage units 24b1 to 24bn are read from the storage unit 21. Then, the propagation unit 23 propagates the read weighting coefficients to the ΔE calculation circuit using the weighting coefficients, and updates the weighting coefficients stored in the storage units of the ΔE calculation circuit using the newly read weighting coefficients.
[0196] In addition, the updating unit 26 is based on the index = p, q, r, s, g p and g s To update the four state variables stored in the storage unit 26a (step S9).
[0197] The control unit 28 determines whether the number of times of the status update processing in the above steps S2 to S9 has reached the predetermined number of times N1 (step S10). In the case where the number of times of the status update processing has not reached the predetermined number of times N1, the processing from step S2 is repeated.
[0198] When the number of times of the state update processing has reached the predetermined number of times N1, the control unit 28 determines whether the number of changes in T (the number of changes in temperature) has reached the predetermined number of times N2 (step S11).
[0199] If the number of temperature changes has not reached the predetermined number of times N2, the control unit 28 changes T (lowers the temperature) (step S12). The manner of changing the predetermined number of times N1, N2, and the value of T (how much to reduce at a time, etc.) is determined based on a predetermined temperature change schedule, etc. After the processing in step S12, the processing from step S2 is repeated.
[0200] For example, when the number of changes in temperature has reached the predetermined number of times N2, the control unit 28 outputs the values of all state variables stored in the storage unit 26a at this time to the control device 30 as a calculation result (step S13), and terminates the processing. Note that the control unit 28 may calculate energy based on the values of all state variables each time a state transition occurs, sequentially update the values of all state variables in which the minimum energy is obtained, and output the values of all state variables as a solution when the number of changes in temperature has reached the predetermined number of times N2.
[0201] Note that the order of the above processing is not limited to the above example, and the order may be changed as appropriate.
[0202] According to the optimization device 20 of the second embodiment as described above, effects similar to those of the optimization device 10 of the first embodiment can be obtained. That is, the calculation time of the optimization problem with bidirectional one-hot constraints can be shortened, and the hardware for storing weighting coefficients can be reduced.
[0203] In addition, the above reading W il and W jk The efficient technique can reduce the number of accesses to the storage unit 21 and can be expected to further reduce the computation time.
[0204] Fig.17 : is a figure which shows the simulation result which shows the calculation shortening effect in the case of using the optimization apparatus of the second embodiment.
[0205] The combinatorial optimization problem to be calculated is "2QAP 1K had 12" which is one of the quadratic assignment problems, and the correct answer is known. The horizontal axis represents the number of iterations (the number of update processes as described above), and the vertical axis represents the number of copies that have obtained the correct answer (the number of optimization devices 20).
[0206] Fig.17 The result 50 in the case of using the 4-bit conversion optimization device 20 as described above and the result 51 in the case of using the 2-bit conversion optimization device that performs the search and excluding the state that satisfies the one-hot constraint in one direction of the column or row are shown. In addition, the result 52 in the case of using the conventional 1-bit conversion optimization device is shown.
[0207] like Fig.17 As shown, compared with the 2-bit conversion optimization device, the optimization device 20 can reduce the number of iterations to obtain the correct answer to 1 / 100 or less. For example, when the number of copies to obtain the correct answer = 15, the number of iterations is reduced to 1 / 337, which indicates that the calculation speed is faster.
[0208] (Third Embodiment)
[0209] The optimization device 20 of the second embodiment is provided with the same number of ΔE calculation circuits 24a1 to 24an as the number of state variables (n=N 2 ). However, the number of ΔE calculation circuits can be less than N 2 Hereinafter, a case where the number of ΔE calculation circuits is less than N will be described. 2 Example of an optimized device.
[0210] Fig.18 : is a diagram showing an example of an optimization device according to the third embodiment. Fig.18 In, with Figure 3 The same elements as those shown in FIG. 1 are denoted by the same reference numerals, and description thereof is omitted.
[0211] In the optimization device 60 of the third embodiment, the propagation unit 61 is different from Figure 3 The number of the propagation unit 23 in the ΔE calculation circuits 62a1 to 62aN is N. In addition, the optimization device 60 includes a propagation control unit 63 and n (=N 2 ) ΔE storage units 64a1 to 64an.
[0212] The number of circuits in which ΔE is calculated is equal to N 2 In the optimization device 20 of the second embodiment, W il and W jkIn contrast, in the case where the number of ΔE calculation circuits is less than N, 2 In the optimization device 60 of the third embodiment, W il and W jk With N 2 Each of the h updating circuits 22a1 to 22an stores accordingly. For example, in the case where j=1 in equation (3), W to be used for calculating ΔE1 il and W jk (For example, in Figure 5 After the state transition in Figure 6 W 4,21 and W 1,24 ) is stored in association with the h update circuit 22a1 that updates h1.
[0213] However, in the following, in order to simplify the circuit configuration, W il and W jk Each h update circuit that updates a local field represented by index = l and k is stored correspondingly. Figure 5 After the state transition shown, with the update h 21 The h update circuit correspondingly stores W as the value to be used to calculate ΔE1 il W 4,21 , and update h 24 The h update circuit correspondingly stores W as the value to be used to calculate ΔE1 jk W 1,24 .
[0214] In the optimization device 60, the propagation unit 61 includes storage units 61a1 to 61an respectively provided corresponding to the h update circuits 22a1 to 22an. When the corresponding switch among the switches 21b1 to 21bn is turned on, each of the storage units 61a1 to 61an stores the weighting coefficient read from the storage unit 21. For example, when the switch 21b1 is turned on, the weighting coefficient read from the storage unit 21 is 11 To W n1 When the switch 21bn is turned on, the W read from the storage unit 21 1n To W nn Any of the above is stored in the storage unit 61an. The storage units 61a1 to 61an can be implemented by, for example, an electronic circuit such as a register or an SRAM.
[0215] Furthermore, the propagation control unit 61b is different from the propagation control unit 23c of the optimization device 20 of the first embodiment.
[0216] The propagation control unit 61b serves to distribute the n local fields (or the weighting coefficients or state variables corresponding thereto) to the N ΔE calculation circuits 62a1 to 62aN. 2 :N multiplexer, N 2 A :N multiplexer is implemented by multiple N:1 multiplexers.
[0217] At the same time, the propagation control unit 63 functions to distribute the energy changes calculated by the N ΔE calculation circuits 62a1 to 62aN to the N:N ΔE storage units 64a1 to 64an. 2 Demultiplexer. N:N 2 The demultiplexer is implemented by multiple 1:N demultiplexers.
[0218] Hereinafter, a case where N=5 will be described as an example.
[0219] Fig.19 is a diagram showing an example of a propagation control unit that propagates a local field and a weighting coefficient.
[0220] The propagation control unit 61 b includes twenty-five switches 70 a 1 to 70 a 25 , multiplexers 71 a 1 , 71 a 2 , 71 a 3 , 71 a 4 , 71 a 5 , 71 b 1 , 71 b 2 , 71 b 3 , 71 b 4 , and 71 b 5 , and a control signal generating circuit 72 .
[0221] Each end of the switches 70a1 to 70a25 receives a signal from h1 to h2. 25 As an input, the other end of the switch 70a1 is connected to one of the five input terminals of any two of the multiplexers 71a1 to 71a5, except for the switches 70a1 and 70a25. The other end of the switch 70a1 is connected to one of the five input terminals of the multiplexer 71a1, and the other end of the switch 70a25 is connected to one of the five input terminals of the multiplexer 71a5. Note that in the case where the weighting coefficients are stored in the storage units 61a1 to 61an, the weighting coefficients are also input to the storage units h1 to h2. 25 The switches 70a1 to 70a25 are turned on and off by the control signals cnta1 to cnta25 generated by the control signal generating circuit 72.
[0222] Each of the multiplexers 71a1 to 71a5 selects and outputs one of the local fields (or weighting coefficients) input to the five input terminals according to any of the control signals cntb1, cntb2, cntb3, cntb4, and cntb5 output by the control signal generation circuit 72.
[0223] The local fields (or weighting coefficients) output by the multiplexers 71a1 to 71a5 are input to the five input terminals of the multiplexers 71b1 to 71b5, respectively. Each of the multiplexers 71b1 to 71b5 selects and outputs one of the local fields (or weighting coefficients) input to the five input terminals according to any of the control signals cntc1, cntc2, cntc3, cntc4, and cntc5 output by the control signal generating circuit 72.
[0224] The control signal generation circuit 72 is based on the index = p, q, r, s, g p or s Generate and output control signals cnta1 to cnta25, cntb1 to cntb5, and cntc1 to cntc5. The control signal generation circuit 72 may include, for example, a memory for storing the thermal level management table as described above. In this case, the control signal generation circuit 72 generates each of the above control signals based on the thermal level management table. Based on index = p, q, r, s, g p and g s Update the heat management table.
[0225] Note that, for example, the propagation h in equation (3) j In the cycle, the x1 to x2 input to the propagation control unit 61b are n With h j Let’s spread it together.
[0226] In the case of N = 5, Fig.19 As shown, five ΔE calculation circuits 62a1, 62a2, 62a3, 62a4 and 62a5 are provided.
[0227] The local fields (or weighting coefficients or state variables) are supplied from the multiplexer 71b1 to the ΔE calculation circuit 62a1, from the multiplexer 71b2 to the ΔE calculation circuit 62a2, and from the multiplexer 71b3 to the ΔE calculation circuit 62a3, respectively. In addition, the local fields (or weighting coefficients) are supplied from the multiplexer 71b4 to the ΔE calculation circuit 62a4, and from the multiplexer 71b5 to the ΔE calculation circuit 62a5, respectively.
[0228] Notice, Fig.19 A circuit configuration of an example of the ΔE calculation circuit 62a1 is shown. The ΔE calculation circuits 62a2 to 62a5 have similar circuit configurations.
[0229] The ΔE calculation circuit 62a1 includes a multiplier 75, an adder 76, a storage unit 77, a selector 78, and a control signal generation circuit 79.
[0230] The multiplier 75 multiplies the h output by the multiplexer 71b1 byi 、h j 、h k 、h l , W il and W jk The code (+1 or -1) outputted from the control signal generating circuit 79 is multiplied.
[0231] The adder 76 outputs an addition result of the multiplication result output by the multiplier 75 and the value stored in the storage unit 77 .
[0232] The storage unit 77 stores the addition result output by the adder 76. The storage unit 77 can be realized by, for example, an electronic circuit such as a register or an SRAM.
[0233] The selector 78 selects and outputs a predetermined positive value (ΔEmax) stored in a register (not shown) or the energy change (ΔEmax) stored in the storage unit 77 based on a selection signal output by the control signal generating circuit 79, for example. j ).
[0234] The control signal generation circuit 79 is for the propagation control unit 61b to propagate h i Each cycle of the same i 、h j 、h k 、h l , W il and W jk In addition, in the x output by the multiplexer 71b1 j When x is 1, the control signal generating circuit 79 outputs a selection signal for selecting ΔEmax to the selector 78, and j When ΔE is 0, the control signal generation circuit 79 is used to select j The selection signal is output to the selector 78.
[0235] Notice, Figure 2 The ΔE calculation circuits 24a1 to 24an in FIG. 2 can also be realized by a circuit configuration almost similar to that described above.
[0236] Use as Fig.19 The configuration of the five multiplexers 71a1 to 71a5 in corresponds to the use of five buses. 2 Each energy change requires 4N cycles. In the case of N=5, 20 cycles are required.
[0237] Fig. 20 is a diagram showing an example of energy change calculated in each cycle in the case of N=5.
[0238] In cycles 1 to 4, calculate ΔE for j=1 to 5 j In cycle 1, h i is propagated by the propagation control unit 61b, and the control signal generation circuit 79 outputs +1 as the code. Therefore, ΔE stored in the storage unit 77 j h i In cycle 2, h j and x j is propagated by the propagation control unit 61b, and the control signal generation circuit 79 outputs -1 as the code. Therefore, ΔE is calculated. j -=h j , and the ΔE stored in the storage unit 77 j h i -h j In cycle 3, h k and W jk is propagated by the propagation control unit 61b, and the control signal generation circuit 79 outputs -1 as the code. Therefore, ΔE is calculated. j -=h k +W jk , and the ΔE stored in the storage unit 77 j h i -h j -h k -W jk In cycle 4, h l and W il is propagated by the propagation control unit 61b, and the control signal generation circuit 79 outputs the signal for h l +1 for W il = -1 as the code. Therefore, ΔE is calculated j + = h l -W il , and the ΔE stored in the storage unit 77 j h i -h j -h k +h l –W jk -W il , that is, ΔE expressed by equation (3) j Note that the calculation is similar to the case where x has a value of 1. j The corresponding ΔE j The ΔE calculation circuit outputs ΔEmax instead of ΔE j .
[0239] In cycles 5 to 8, the same process is performed for j=6 to 10, and in cycles 9 to 12, the same process is performed for j=11 to 15. Furthermore, in cycles 13 to 16, the same process is performed for j=16 to 20, and in cycles 17 to 20, the same process is performed for j=21 to 25.
[0240] Next, the propagation control unit 61b will be described as follows. Figure 5 An example of operation in each of the above-described cycles after the state transition is shown.
[0241] Fig.21 is a diagram showing an example of the operation of the propagation control unit in cycle 1. Note that Fig.21 The dashed line shows the i A signal line that is invalid when propagating through it.
[0242] exist Figure 5 After the state transition shown in , x4 among x1 to x5 becomes 1. Therefore, h4 is propagated to the ΔE calculation circuits 62a1 to 62a5 as h1. In this case, the switch 70a4 among the switches 70a1 to 70a25 is turned on, and the other switches are turned off. Then, the multiplexer 71a1 selects and outputs the input h4, and the multiplexers 71b1 to 71b5 select and output h4 as the output of the multiplexer 71a1. Thus, ΔE1 to ΔE5 calculated by the ΔE calculation circuits 62a1 to 62a5 in cycle 1 become h4.
[0243] Fig. 22 is a diagram showing an example of the operation of the propagation control unit in cycle 2. Note that Fig. 22 The dashed line shows the j and x j A signal line that is invalid when propagating through it.
[0244] In cycle 2, h1 to h5 and x1 to x5 are propagated. In this case, switches 70a1 to 70a5 among switches 70a1 to 70a25 are turned on, and the other switches are turned off. Then, multiplexer 71a1 selects and outputs inputs h1 and x1, multiplexer 71a2 selects and outputs inputs h2 and x2, and multiplexer 71a3 selects and outputs inputs h3 and x3. In addition, multiplexer 71a4 selects and outputs inputs h4 and x4, and multiplexer 71a5 selects and outputs inputs h5 and x5.
[0245] Furthermore, the multiplexer 71b1 selects and outputs h1 and x1 as the output of the multiplexer 71a1, and the multiplexer 71b2 selects and outputs h2 and x2 as the output of the multiplexer 71a2. The multiplexer 71b3 selects and outputs h3 and x3 as the output of the multiplexer 71a3, and the multiplexer 71b4 selects and outputs h4 and x4 as the output of the multiplexer 71a4. The multiplexer 71b5 selects and outputs h5 and x5 as the output of the multiplexer 71a5.
[0246] Thus, the ΔE calculation circuit 62a1 calculates ΔE1=h1, the ΔE calculation circuit 62a2 calculates ΔE2=h2, the ΔE calculation circuit 62a3 calculates ΔE3=h3, and the ΔE calculation circuit 62a4 calculates ΔE4=h4, and the ΔE calculation circuit 62a5 calculates ΔE5=h5.
[0247] Fig.23 is a diagram showing an example of the operation of the propagation control unit in cycle 3. Note that Fig.23 In the figure, the dashed line indicates the k and W jk A signal line that is invalid when propagating through it.
[0248] In cycle 3, the propagation is h k h4, h9, h 14 、h 19 and h 24 As W jk W 4,4 , W 3,9 , W 5,14 , W 2,19 and W 1,24 .exist Figure 5 After the state transition shown in FIG. 1 , the address decoder 27 reads the weighting coefficients from the storage unit 21 and stores the weighting coefficients in h4, h9, h10 and h11 corresponding to the update storage units 61a1 to 61an. 14 、h 19 and h 24 The h updates the storage unit corresponding to the circuit.
[0249] In this case, switches 70a4, 70a9, 70a14, 70a19, and 70a24 among switches 70a1 to 70a25 are turned on, and the other switches are turned off. Then, the multiplexer 71a1 selects and outputs the inputs h4 and W 4,4 , multiplexer 71a2 selects and outputs input h9 and W 3,9 , and the multiplexer 71a3 selects and outputs the input h 14 and W 5,14In addition, the multiplexer 71a4 selects and outputs the input h 19 and W 2.19 , and the multiplexer 71a5 selects and outputs the input h 24 and W 1,24 .
[0250] In addition, the multiplexer 71b1 selects and outputs h as the output of the multiplexer 71a5. 24 and W 1,24 , and the multiplexer 71b2 selects and outputs h as the output of the multiplexer 71a4 19 and W 2,19 The multiplexer 71b3 selects and outputs h9 and W as the output of the multiplexer 71a2. 3,9 , and the multiplexer 71b4 selects and outputs h4 and W as the output of the multiplexer 71a1 4,4 The multiplexer 71b5 selects and outputs h as the output of the multiplexer 71a3. 14 and W 5,14 .
[0251] Thus, the ΔE calculation circuit 62a1 calculates ΔE1-=h 24 +W 1,24 , ΔE calculation circuit 62a2 calculates ΔE2-=h 19 +W 2,19 , and the ΔE calculation circuit 62a3 calculates ΔE3-=h9+W 3,9 In addition, the ΔE calculation circuit 62a4 calculates ΔE4-=h4+W 4,4 , and the ΔE calculation circuit 62a5 calculates ΔE5-=h 14 +W 5,14 .
[0252] Fig.24 is a diagram showing an example of the operation of the propagation control unit in cycle 4. Note that Fig.24 In the figure, the dashed line indicates the l and W il A signal line that is invalid when propagating through it.
[0253] In cycle 4, the propagation is h l h4, h8, h 15 、h 17 and h 21 As W il W 4,4 , W 4,8 , W 4,15 , W 4,17 and W 4,21 .exist Figure 5After the state transition shown in FIG. 1 , the address decoder 27 reads the weighting coefficients from the storage unit 21 and stores the weighting coefficients in h4, h8, h9 and h10 corresponding to the update storage units 61a1 to 61an. 15 、h 17 and h 21 The h updates the storage unit corresponding to the circuit.
[0254] In this case, switches 70a4, 70a8, 70a15, 70a17, and 70a21 among switches 70a1 to 70a25 are turned on, and the other switches are turned off. Then, the multiplexer 71a1 selects and outputs the inputs h4 and W 4,4 , multiplexer 71a2 selects and outputs input h8 and W 4,8 , and the multiplexer 71a3 selects and outputs the input h 15 and W 4,15 In addition, the multiplexer 71a4 selects and outputs the input h 17 and W 4,17 , and the multiplexer 71a5 selects and outputs the input h 21 and W 4,21 .
[0255] In addition, the multiplexer 71b1 selects and outputs h as the output of the multiplexer 71a5. 21 and W 4,21 , and the multiplexer 71b2 selects and outputs h as the output of the multiplexer 71a4 17 and W 4,17 The multiplexer 71b3 selects and outputs h8 and W as the output of the multiplexer 71a2. 4,8 , and the multiplexer 71b4 selects and outputs h4 and W as the output of the multiplexer 71a1 4,4 The multiplexer 71b5 selects and outputs h as the output of the multiplexer 71a3. 15 and W 4,15 .
[0256] Thus, the ΔE calculation circuit 62a1 calculates ΔE1+=h 21 -W 4,21 , ΔE calculation circuit 62a2 calculates ΔE2+=h 17 -W 4,17 , and the ΔE calculation circuit 62a3 calculates ΔE3+=h8-W 4,8 In addition, the ΔE calculation circuit 62a4 calculates ΔE4+=h4-W 4,4 , and the ΔE calculation circuit 62a5 calculates ΔE5+=h 15 -W 4,15 .
[0257] The calculation of equation (3) for ΔE1 to ΔE5 is completed in the above cycles 1 to 4. Note that since x4=1, ΔEmax is output as ΔE4 instead of the value calculated as described above.
[0258] In the subsequent cycles, ΔE6 to ΔE are calculated by similar control. 25 .
[0259] Fig.25 is a diagram showing an example of a propagation control unit in which propagation energy varies.
[0260] The propagation control unit 63 propagates the energy changes calculated by the five ΔE calculation circuits 62a1 to 62a5 to any five of the twenty-five ΔE storage units 64a1 to 64a25. The propagation control unit 63 includes twenty-five switches 63a1 to 63a25 and a switch control circuit 63b. Note that Fig.25 The five black circles in the figure act as demultiplexers.
[0261] The switch control circuit 63b turns on any five of the switches 63a1 to 63a25 and turns off the other switches to propagate the energy changes calculated by the ΔE calculation circuits 62a1 to 62a5 to any five of the ΔE holding units 64a1 to 64a25.
[0262] For example, Fig. 20 As shown, ΔE1 to ΔE are calculated in twenty cycles. 25 In the case of , in cycle 4 when calculation of ΔE1 to ΔE5 is completed, five switches of the switches 63a1 to 63a25 connected to the ΔE storage unit storing ΔE1 to ΔE5 are turned on. 10 In cycle 8 of the calculation, switches 63a1 to 63a25 are connected to store ΔE6 to ΔE 10 The five switches of the ΔE saving unit are turned on. 11 To ΔE 15 In cycle 12 of the calculation, switches 63a1 to 63a25 are connected to store ΔE 11 To ΔE 15 The five switches of the ΔE saving unit are turned on. 16 To ΔE 20 In cycle 16 of the calculation, the switches 63a1 to 63a25 are connected to store ΔE 16 To ΔE 20 The five switches of the ΔE saving unit are turned on. 21 To ΔE 25 In the calculation cycle 20, the switches 63a1 to 63a25 are connected to save ΔE21 To ΔE 25 The five switches of the ΔE saving unit are turned on.
[0263] Therefore, we can obtain Figure 2 The n=25 energy changes are input to the conversion selection unit 25. Other than that, operations similar to those of the optimization device 20 of the second embodiment are performed. Note that the conversion selection unit 25 inputs N energy changes at a time. 2 In the case of N energy changes and selecting one energy change from N energy changes, it is not necessary to Fig.18 The propagation control unit 63 and ΔE holding units 64a1 to 64an are shown.
[0264] As described above, even if the number of ΔE calculation circuits 62a1 to 62aN is less than N 2 , by propagating the local fields and weighting coefficients to be used for calculation through the propagation unit 61, an effect similar to that of the optimization device 20 of the second embodiment can also be obtained.
[0265] (Revise)
[0266] By the way, except for N which is subject to bidirectional one-hot constraints 2 In addition to state variables, the state variables included in the evaluation function may also include state variables that are not subject to bidirectional one-hot constraints.
[0267] In this case, the ΔE calculation circuit may be configured as follows.
[0268] Fig.26 is a diagram showing a modification of the ΔE calculation circuit. Fig.26 In, with Fig.19 Similar elements to the ΔE calculation circuit 62 a 1 shown in FIG. 6 are denoted by the same reference numerals.
[0269] The ΔE calculation circuit 80 in the modification is different from that in the control signal generating circuit 79a. Fig.19 The control signal generating circuit 79 shown in FIG. 2A and 2B are further shown. In addition, it is assumed that each state variable and a flag value indicating whether the state variable is subject to the bidirectional one-hot constraint are stored in a storage unit (not shown).
[0270] For example, in x j is a state variable subject to a bidirectional one-hot constraint, the flag value (f j ) is 1. On the contrary, at x j In the case where the state variable is not subject to the bidirectional one-hot constraint, the flag value (f j ) is 0.
[0271] The control signal generating circuit 79a reads f j and xj , and at f j =1, +1 or -1 is output as a control signal to the h as described above. i 、h j 、h k 、h l , W il and W jk . In f j =0, the control signal generating circuit 79a outputs 0 as the control signal to h i 、h j 、h k 、h l , W il and W jk , and 2x j -1 output to h j Therefore, ΔE j =(2x j -1)h j . 2x j -1 is x j The change in (2x j -1)h j This means that among all state variables, only x j The energy changes under changing conditions. Note that in f j =0, the control signal generating circuit 79a causes the selector 78 to select ΔE j .
[0272] Note that in the conversion selection unit 25, j =0 j Select ΔE in case of change j In the case of j Then, the local field is also updated according to the update.
[0273] For example, with this configuration, even if the state variables included in the evaluation function include N other than those subject to the bidirectional one-hot constraint 2 It also becomes possible to process state variables other than the state variables that are not subject to the bidirectional one-hot constraint.
[0274] In the above, one aspect of the optimization device and the optimization method of the present embodiment has been described based on the embodiment; however, these are only examples and are not limited to the above description.
[0275] [reference numerals list]
[0276] 10 Optimize the device
[0277] 11, 13bi, 13bj, 15a storage unit
[0278] 12 Communication Units
[0279] 13 Computational Unit
[0280] 13ai to 13aj ΔE calculation circuit
[0281] 14 Conversion selection unit
[0282] 15 Update Unit
[0283] 16 Address decoder
[0284] [Citation list]
[0285] Japanese Patent Publication No. 7-36473.
[0286] Japanese Patent Publication No. 2003-263890.
Claims
1. An optimization device for searching for a solution to shorten the computation time of an optimization problem with a bidirectional one-hot constraint, the optimization device comprising: a computing unit configured to compute, for each of a plurality of first state variables, a first change in the energy of an Ising model representing the optimization problem, the first change being a change generated by changing the values of four state variables among the plurality of state variables of the Ising model to satisfy the bidirectional one-hot constraint, each of the plurality of first state variables being a state variable having a value of 0 before the change among the plurality of state variables in the Ising model, the computing unit configured to compute the first change to satisfy the bidirectional one-hot constraint, and in a case where N square state variables included in an evaluation function representing the energy of the Ising model are arranged in N rows and N columns, a sum of the values of the state variables included in each row is 1 and a sum of the values of the state variables included in each column is 1, wherein N is an integer of 2 or greater; a conversion selection unit configured to select four state variables allowed to be changed based on the first change amount calculated for each of the plurality of first state variables, and output identification information for identifying the selected four state variables; and An updating unit is configured to update the values of the four state variables that are allowed to change based on the identification information.
2. The optimization device according to claim 1, wherein: The calculation unit includes a plurality of energy change calculation circuits, each of which is configured to calculate any of the first change amounts for a corresponding plurality of first state variables. The optimization device also includes: A propagation unit is configured to propagate any of N square local fields to each of the multiple energy change calculation circuits based on the identification information, and the N square local fields each represent the size of a second change in the energy of the Ising model caused by a change in one of the N square state variables, and the local field will be used to calculate the first change.
3. The optimization device according to claim 2, wherein: The propagation unit is configured to provide, based on the identification information, to each of the plurality of energy change calculation circuits a first local field for any one of the plurality of first state variables, a second local field for a second state variable included in the same row as the first state variable and having a value of 1, a third local field for a third state variable included in the same column as the first state variable and having a value of 1, and a fourth local field for a fourth state variable included in the same column as the second state variable and in the same row as the third state variable and having a value of 0.
4. The optimization device according to claim 3, wherein: The propagation unit provides each of the N square local fields to any of the plurality of energy change calculation circuits in a time division manner or in parallel using a plurality of buses.
5. The optimization device according to any one of claims 1 to 4, further comprising: a storage unit configured to store a plurality of weighting coefficients, each of which indicates a magnitude of an interaction between every two of the N square state variables; as well as an address decoder configured to specify an address of the storage unit based on the identification information, the storage unit storing a weighting coefficient group to be used for calculating the first change amount among the plurality of weighting coefficients, wherein The storage unit reads the weighting coefficient group at a time based on the address.
6. An optimization method implemented by an optimization device for searching for a solution to shorten the computation time of an optimization problem with a bidirectional one-hot constraint, the optimization method comprising: For each of a plurality of first state variables, calculating a first change in energy of an Ising model representing the optimization problem, the first change being a change generated by changing values of four state variables among a plurality of state variables of the Ising model to satisfy the bidirectional one-hot constraint, wherein each of the plurality of first state variables is a state variable among a plurality of state variables in the Ising model having a value of 0 before the change; Calculating the first variation to satisfy the bidirectional one-heat constraint, in a case where N square state variables included in an evaluation function representing the energy of the Ising model are arranged in N rows and N columns, a sum of values of the state variables included in each row is 1 and a sum of values of the state variables included in each column is 1, wherein N is an integer of 2 or greater; selecting the four state variables allowed to change based on the change amount calculated for each of the plurality of first state variables, and outputting identification information for identifying the selected four state variables; and The values of the four state variables that are allowed to change are updated based on the identification information.
