Optimized learning method based on memory model

Abstract

The invention discloses an optimization learning method based on a memory model, and the method comprises the steps: constructing a long memory learning model based on fractional differential; constructing a short memory model based on the long memory learning model and the attenuation characteristics; discretizing the short memory model to obtain a discrete memory model; and based on a particle swarm optimization algorithm and the discrete memory model, optimizing the memory parameters, and obtaining a memory optimization model. According to the method and the device, the memory learning of people is expanded to machine learning on the basis of memory characteristics based on the learning of people, so that the machine learning has better convergence and better learning speed at the same time. According to the method, the historical information is weighted according to the memory characteristics by utilizing the memory attenuation characteristics, the weight of the historical information exceeding a certain length is relatively small and can be ignored, and the learning speed and the learning efficiency are both considered.

Description

Technical Field

[0001] This application belongs to the field of deep learning technology, and specifically relates to an optimization learning method based on a memory model. Background Art

[0002] Artificial intelligence is about simulating the human brain with computers, enabling machines to acquire knowledge, learn, judge, and make decisions like humans. Knowledge acquisition and updating are the foundation of judgment and decision-making.

[0003] The human brain's ability to learn stems from its memory function. Without memory, learning would be meaningless. The learning process is a process of knowledge renewal: old knowledge is forgotten and new knowledge is acquired. Clearly, learning is an asymptotic process, with old knowledge gradually fading and new knowledge being acquired. Individuals have varying memory capacities. Those with strong memory abilities have enhanced learning abilities and are more adaptable to new and changing knowledge. For information that changes infrequently, memorization often requires less effort.

[0004] The human brain learns based on its memory function. Machine learning is a computer simulation of human brain learning. Therefore, machine learning should also be based on memory, adjusting memory according to changes in the object being learned. For objects with large changes, memory needs to be strengthened to consolidate learning; for objects with small changes, learning can be weakened to increase learning speed.

[0005] Although machine learning has become a current research hotspot, memory learning has also received attention. However, existing research has rarely considered the changing characteristics of the learning object and adjusted the memory scale for learning. This can lead to difficulties such as different learning convergence speeds and non-convergence for different learning objects. Summary of the Invention

[0006] The present application provides an optimization learning method based on a memory model to solve the technical problems in the prior art of having different learning speeds for different learning objects or even non-convergence of learning.

[0007] To solve the above technical problems, a technical solution adopted in this application is: an optimization learning method based on a memory model, comprising:

[0008] S1. Construct a long memory learning model based on fractional differentiation;

[0009] S2. Construct a short memory model based on the long memory learning model and decay characteristics;

[0010] S3. discretize the short memory model to obtain a discrete memory model;

[0011] S4. Based on the particle swarm optimization algorithm and the discrete memory model, optimize the memory parameters and obtain the memory optimization model.

[0012] Furthermore, the method of step S1 includes:

[0013] Based on formula (1), a long memory learning model is obtained; wherein formula (1) is:

[0014]

[0015] Among them, x(t) represents input information, represents output information, represents the v-th Caputo fractional differential of the state variable x(t) with respect to time t, u(t) is the unit step function, Γ(.) is the gamma function, and the symbol * represents the convolution operation.

[0016] Furthermore, the method of step S2 includes:

[0017] Based on formula (2), a short memory learning model is obtained; wherein formula (2) is:

[0018]

[0019] Wherein, L is the historical information of preset length.

[0020] Furthermore, the method of step S3 includes:

[0021] Based on formula (3), a discrete memory model is obtained; wherein formula (3) is:

[0022] y(n)=k([x(n)-(n-1)]1 -α +[x(n-1)-(n-2)]2 -α +…[x(nM)-(nM-1)](M+1) -α ) (3);

[0023] Where y(n) is the discrete sampling of y(t), x(n) is the discrete sampling of x(t); M is a constant related to L, reflecting the number of historical data related to the memory length, and k is the learning rate constant.

[0024] For further information, see Figure 2 The method of step S4 comprises:

[0025] S41. Initialize the particle swarm and calculate the cost function of each particle;

[0026] S42. Setting memory parameters based on optimal individual differences of the particle swarm;

[0027] S43. Update the individual optimal position and local optimal solution based on the discrete memory model and memory parameters;

[0028] S44. Based on the individual best position, update the global best position;

[0029] S45. Based on the global optimal position and the local optimal solution, update the velocity and position of the particle;

[0030] S46. Determine whether the cost function is less than a preset threshold; if not, repeat the above steps; if so, save the global optimal solution as the model parameter.

[0031] Furthermore, the method of step S41 includes:

[0032] Based on formula (4), the cost function is obtained; wherein, the formula (4) is:

[0033]

[0034] Among them, cost(y) is the cost function, represents the output information sampling value obtained in the experiment, y(n) is the discrete sampling of y(t) obtained by solving formula (2)

[0035] Furthermore, the method of step S44 includes:

[0036] Based on formula (5), the optimal position of the individual is obtained; where formula (5) is:

[0037] pbest i,j =p i-1,j +k([p i,j -p i-1,j ]1 -α +[p i-1,j -p i-2,j ]2 -α +…[p i-M,j -p i-M-1,j ](M+1) -α ) (5);

[0038] Among them, pbest i,j represents the local optimal value of the jth particle in the i-th generation; p i,j Indicates the particle position.

[0039] The beneficial effects of this application are as follows: This application is based on the memory characteristics of human learning and extends human memory learning to machine learning, so that machine learning can converge better while having a better learning speed. This application uses the memory decay characteristics to weight historical information according to the memory characteristics. Historical information exceeding a certain length has a smaller weight and can be ignored, taking into account both learning speed and learning efficiency. This application dynamically adjusts the memory decay characteristics and memory length based on the characteristics of signal changes, which is closer to human learning functions and makes learning more intelligent. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of an embodiment of an optimization learning method based on a memory model of the present application;

[0041] Figure 2 yes Figure 1 A flow chart of an embodiment of step S4 in FIG.

[0042] Figure 3 This is an optimized search result graph of an embodiment of the memory model-based optimization learning method of the present application. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to specific embodiments.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from the description. Therefore, the present invention is not limited to the specific embodiments disclosed in the following specification.

[0045] See Figure 1 , Figure 1 1 is a flow chart of an embodiment of the memory model-based optimization learning method of the present application. The method includes:

[0046] S1. Construct a long memory learning model based on fractional-order differentiation.

[0047] Specifically, based on formula (1), a long memory learning model is obtained; wherein formula (1) is:

[0048]

[0049] Among them, x(t) represents input information, represents output information, represents the v-th Caputo fractional differential of the state variable x(t) with respect to time t, u(t) is the unit step function, Γ(.) is the gamma function, and the symbol * represents the convolution operation.

[0050] The fractional order model shows that the output has a memory characteristic for changes in historical input information. It has a memory characteristic and can well reflect the memory characteristic of neurons for historical information. The closer the information is to the current one, the greater the weight.

[0051] S2. Construct a short memory model based on the long memory learning model and decay characteristics.

[0052] Specifically, the long memory model is a convolutional model, and the amount of computation increases exponentially over time. This increased computational overhead also increases computational latency. The memory weight gradually decreases over time, and the effect of historical information beyond a certain length can be ignored. To balance learning convergence and learning speed, a short memory learning model can be obtained by truncating historical information of a certain length, L, for learning, as shown below:

[0053] Based on formula (2), a short memory learning model is obtained; wherein formula (2) is:

[0054]

[0055] Wherein, L is the historical information of preset length.

[0056] The differential order v is an important parameter of the memory scale. The smaller v is, the slower the effect of historical information decays, and a longer memory length is required. Conversely, the larger v is, the faster the effect of historical information decays, and the role of current information becomes more prominent, and the required memory length can be appropriately shortened. In particular, when v = 1, the fractional order differential becomes an integer order differential. It is an ideal impulse response model, and the memory length of historical information is 0.

[0057] Obviously, when the information changes little, the integer-order differential model can simplify the calculation, but when the information changes greatly, it cannot weaken the noise and cannot optimize the learning speed.

[0058] S3. Discretize the short memory model to obtain a discrete memory model.

[0059] Specifically, based on formula (3), a discrete memory model is obtained; wherein formula (3) is:

[0060] y(n)=k([x(n)-(n-1)]1 -α +[x(n-1)-(n-2)]2 -α +…[x(nM)-(nM-1)](M+1) -α ) (3);

[0061] Where y(n) is the discrete sampling of y(t), x(n) is the discrete sampling of x(t); M is a constant related to L, reflecting the number of historical data related to the memory length, and k is the learning rate constant.

[0062] S4. Based on the particle swarm optimization algorithm and the discrete memory model, optimize the memory parameters and obtain the memory optimization model.

[0063] Specifically, step S4 includes:

[0064] S41. Initialize the particle swarm and calculate the cost function of each particle.

[0065] Specifically, set the size N of the particle swarm, the initial position (x 1,0 ,x 2,0 ,…,x N,0 ), initial velocity (v 1,0 ,v 2,0 ,…,v N,0 ); Calculate the cost function f(x i,j ), f(x i,j ) represents the cost function of the i-th particle in the j-th generation.

[0066] Based on formula (4), the cost function is calculated; the specific formula is as follows:

[0067]

[0068] cost(y) is the cost function, represents the output information sampling value obtained in the experiment, and y(n) is the discrete sampling of y(t) calculated according to formula (2).

[0069] S42. Set memory parameters based on the optimal individual difference changes of the particle swarm.

[0070] Specifically, the memory length and memory order α are adjusted based on the difference between the current optimal individual and the previous optimal individual. If the difference is large, the memory length is increased and α is decreased to increase the memory effect and smooth out the large difference. If the difference is small, the memory length is decreased and α is increased to weaken the memory effect and increase the learning speed.

[0071] S43. Update the individual optimal position and local optimal solution based on the discrete memory model and memory parameters.

[0072] Specifically, combined with the memory characteristics, the individual optimal position is updated according to formula (5):

[0073] pbest i,j =p i-1,j +k([p i,j -p i-1,j ]1 -α +[p i-1,j -p i-2,j ]2 -α +…[pi-M,j -p i-M-1,j ](M+1) -α ) (5);

[0074] Among them, pbest i,j represents the local optimal value of the jth particle in the i-th generation; p i,j Indicates the particle position.

[0075] That is, if a current individual is found to be superior to the best historical individual, it is not directly replaced with the current individual. Instead, the best individuals from each generation are weighted according to the memory decay characteristics to determine the current best individual. Even if a particular generation achieves the best individual due to anomalies, the memory weighting can still smooth it out. Even if individual fluctuations occur, these fluctuations can be smoothed through the memory model.

[0076] S44. Update the global optimal position based on the individual optimal position.

[0077] Specifically, the cost functions corresponding to all particles in the current era are compared with the current optimal particle cost function. If a particle cost function is better than the current cost function, the optimal particle position is updated. i represents the global optimal value of the i-th generation.

[0078] S45. Update the particle's velocity and position based on the global optimal position and the local optimal solution.

[0079] Specifically, the particle speed and position are updated by combining the memory characteristics and using the historical global optimal solution and the local optimal solution:

[0080] v i,j =wv i,j +c1r1([pbest i,j -p i,j ]1 -α +[pbest i-1,j -p i-1,j ]2 -α +…[pbest i-M,j -p i-M,j ](M+1) -α )+c2r2([gbest i,j -p i,j ]1 -α +[gbest i-1,j -p i-1,j ]2 -α +…[gbest i-M,j -p i-M,j ](M+1) -α ) (6);

[0081] Where w is the weight, constants c1 and c2 satisfy: c1 + c2 = 1. r1 and r2 are random numbers in the range (0, 1).

[0082] Update particle position: p i+1,j =p i,j +v i,j+1 .

[0083] This particle velocity update updates particle position information based on historical global and local optima. This historical information is weighted according to memory characteristics, with greater weight given to information closer to the current location. This prevents accidental errors from causing search oscillation and trapping in local optima. It also adjusts the memory scale and length based on errors, balancing search speed.

[0084] S46. Determine whether the cost function is less than a preset threshold; if not, repeat the above steps; if so, save the global optimal solution as the model parameters. Using the model parameters, construct a memory optimization model.

[0085] See Figure 3 , search function z = 10 + 3 (x - 3) 2 +2(y-1) 2 The minimum value of -5cos(2(x-2))-8cos(3(y-2)) can be obtained according to the above optimization search steps. When x=2.2.75 and y=1.9472, the optimal solution is: z=1.1923.

[0086] This application is based on the memory characteristics of human learning. It extends human memory learning to machine learning so that machine learning can converge better while having a better learning speed. This application uses the memory decay characteristics to weight historical information according to the memory characteristics. The weight of historical information exceeding a certain length is small and can be ignored, taking into account both learning speed and learning efficiency. This application dynamically adjusts the memory decay characteristics and memory length according to the signal change characteristics, which is closer to the human learning function and makes learning more intelligent.

[0087] The above description is merely an embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An optimization learning method based on a memory model, characterized in that: The following steps are included: S1. Construct a long memory learning model based on fractional differentiation; S2. Constructing a short memory model based on the long memory learning model and the decay characteristics; S3. Discretize the short memory model to obtain a discrete memory model; S4. Based on the particle swarm optimization algorithm and the discrete memory model, optimize the memory parameters and obtain a memory optimization model.

2. The method according to claim 1, characterized in that The method of step S1 comprises: Based on formula (1), the long memory learning model is obtained; wherein, the formula (1) is: Among them, x(t) represents input information, represents output information, represents the v-th Caputo fractional differential of the state variable x(t) with respect to time t, u(t) is the unit step function, Γ(.) is the gamma function, and the symbol * represents the convolution operation.

3. The method according to claim 2, characterized in that The method of step S2 comprises: Based on formula (2), the short memory learning model is obtained; wherein, formula (2) is: Wherein, L is the historical information of preset length.

4. The method according to claim 3, characterized in that The method of step S3 comprises: Based on formula (3), the discrete memory model is obtained; wherein, the formula (3) is: y(n)=k([x(n)-(n-1)]1 -α +[x(n-1)-(n-2)]2 -α +…[x(n-M)-(n-M-1)](M+1) -α ) (3); Where y(n) is the discrete sampling of y(t), x(n) is the discrete sampling of x(t); M is a constant related to L, reflecting the number of historical data related to the memory length, and k is the learning rate constant.

5. The method according to claim 4, characterized in that The method of step S4 comprises: S41. Initialize the particle swarm and calculate the cost function of each particle; S42. Setting memory parameters based on the optimal individual differences of the particle swarm; S43. Based on the discrete memory model and the memory parameters, updating the individual optimal position and the local optimal solution; S44. Based on the individual optimal position, update the global optimal position; S45. Based on the global optimal position and the local optimal solution, update the speed and position of the particle; S46. Determine whether the cost function is less than a preset threshold; if not, repeat the above steps; if so, save the global optimal solution as a model parameter.

6. The method according to claim 5, characterized in that The method of step S41 includes: Based on formula (4), the cost function is obtained; wherein, the formula (4) is: Among them, cost(y) is the cost function, represents the output information sampling value obtained in the experiment, and y(n) is the discrete sampling of y(t) obtained by solving the formula (2).

7. The method according to claim 5, characterized in that The method of step S44 includes: Based on formula (5), the optimal position of the individual is obtained; wherein the formula (5) is: pbest i,j =p i-1,j +k([p i,j -p i-1,j ]1 -α +[p i-1,j -p i-2,j ]2 -α +…[p i-M,j -p i-M-1,j ](M+1) -α ) (5); Among them, pbest i,j represents the local optimal value of the jth particle in the i-th generation; p i,j Indicates the particle position.

Images