A method and system for improving cache performance of online trading platform

Abstract

The embodiment of the present invention discloses a method and system for improving the cache performance of an online trading platform, and belongs to the field of computer technology. The embodiment of the present invention optimizes the cache resource allocation strategy of the online trading platform through a series of steps. First, comprehensive cache performance statistics are obtained to provide a basis for accurate decision-making. The candidate strategies and initial weights are determined using the reinforcement learning decision knowledge learned from the associated platform, and the weights are adjusted with the help of the timeliness requirement level to finally determine the optimal strategy. This process can dynamically adapt to the platform's data access mode and cache state changes, improve the cache hit rate, optimize cache resource utilization, reduce system response delays caused by improper cache strategies, and enhance the overall performance and user experience of the online trading platform in the Internet / supply chain finance scenario.

Description

Technical Field

[0001] The embodiments of the present invention belong to the field of computer technology, and specifically relate to a method and system for improving the cache performance of an online trading platform. Background Art

[0002] With the development of business and the continuous growth of data volume in the field of online trading platforms, cache performance has become a key factor affecting the efficient operation of the platform. Traditional cache strategies are often difficult to adapt to complex and changeable data access patterns and cache resource status. For example, the frequency of data access will change significantly due to factors such as business peaks and valleys, market dynamics, etc., and cache resources are limited. Unreasonable cache strategies may lead to problems such as low cache hit rate, resource waste, and system response delay. Existing technologies lack a method that can dynamically and accurately adjust the cache resource allocation strategy according to the actual situation of the platform. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for improving the cache performance of an online trading platform, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] An embodiment of the present invention provides a method for improving the cache performance of an online trading platform, which is applied to a platform cache performance optimization system. The method includes: obtaining first cache performance statistics of a target online trading platform; wherein the first cache performance statistics include data access state information, cache resource state data, and cache policy response information; determining a candidate cache resource allocation strategy that is paired with first reinforcement learning decision knowledge and the first cache performance statistics, and an initial correlation weight of the candidate cache resource allocation strategy; wherein the first reinforcement learning decision knowledge is obtained by learning associated cache performance statistics of an associated online trading platform through a Q-reinforcement learning algorithm, and the associated online trading platform and the target online trading platform belong to the same online trading scenario; determining an adjustment feature of the candidate cache resource allocation strategy based on the timeliness requirement level of the candidate cache resource allocation strategy; adjusting the initial correlation weight of the candidate cache resource allocation strategy based on the adjustment feature of the candidate cache resource allocation strategy to obtain a correlation adjustment weight of the candidate cache resource allocation strategy; and determining a first current cache resource allocation strategy from the candidate cache resource allocation strategy based on the correlation adjustment weight of the candidate cache resource allocation strategy.

[0005] An embodiment of the present invention provides a platform cache performance optimization system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0006] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] The embodiment of the present invention optimizes the cache resource allocation strategy of the online trading platform through a series of steps. First, comprehensive cache performance statistics are obtained to provide a basis for accurate decision-making. The candidate strategies and initial weights are determined using the reinforcement learning decision knowledge learned from the associated platform, and the weights are adjusted with the help of the timeliness requirement level to finally determine the optimal strategy. This process can dynamically adapt to the platform's data access mode and cache state changes, improve the cache hit rate, optimize cache resource utilization, reduce system response delays caused by improper cache strategies, and enhance the overall performance and user experience of the online trading platform in the Internet / supply chain finance scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The present invention provides a flowchart of a method for improving the cache performance of an online trading platform.

[0009] Figure 2 A schematic diagram of the structure of a platform cache performance optimization system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments in the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present invention.

[0011] The terms "first", "second", etc. in the embodiments of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the embodiments of the present invention can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the embodiments of the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0012] Figure 1 A method for improving the cache performance of an online trading platform is shown, which is applied to a platform cache performance optimization system. The method includes the following steps 200 to 208.

[0013] In online trading platforms under the Internet / supply chain finance scenario, cache performance is crucial to the efficient operation of the platform. The embodiment of the present invention proposes a cache resource allocation strategy determination method based on reinforcement learning, with a platform cache performance optimization system as the execution body, aiming to improve cache performance, thereby improving the operation efficiency of the entire online trading platform.

[0014] Step 200: Obtain first cache performance statistics of a target online trading platform; wherein the first cache performance statistics include data access status information, cache resource status data, and cache policy response information.

[0015] In step 200, the first cache performance statistical data of the target online trading platform obtained includes at least three categories: data access status information, cache resource status data and cache policy response information.

[0016] Data access status information reflects the access situation of various data in the online trading platform. For example, on the supply chain finance platform, it may include the access frequency of data such as credit information, transaction contract information, and fund flow records of different suppliers. For example, in a certain period of time, the access frequency of supplier A's credit information is 100 times per hour, the access frequency of transaction contract information is 50 times per hour, and the access frequency of fund flow records is 80 times per hour. The statistics of these data access frequencies can be updated at certain time intervals (such as 1 hour, 1 day, etc.). At the same time, the time distribution characteristics of data access can also be included. For example, it is found that 10 to 11 am every day is the peak period for access to transaction contract information, while 8 to 9 pm is the trough period for access to fund flow records. This time distribution characteristic is of great significance for the allocation of cache resources.

[0017] The cache resource status data describes the current usage of the cache. For example, the total capacity of the cache is 1000MB, and the currently used cache space is 600MB, and the cache fill rate is 60%. In addition, it can also include the proportion of space occupied by different types of data in the cache. For example, credit information occupies 200MB, transaction contract information occupies 300MB, and fund flow records occupy 100MB. The cache hit rate is also an important cache resource status data. If 600 of the past 1000 data access requests obtained data from the cache, then the cache hit rate is 60%.

[0018] The cache policy response information reflects the response effect of the current cache policy on data access. For example, under a certain cache replacement policy (such as the Least Recently Used (LRU) policy), the change in cache hit rate caused by each replacement operation. For example, after an LRU policy replacement operation, the cache hit rate dropped from 58% to 55%, which indicates that the replacement operation may not be suitable for the current data access pattern. It can also include the effect of the cache prefetching policy. If a certain type of data is prefetched into the cache, observe the proportion of the data accessed in the following period of time after the prefetch. For example, a batch of credit information of new suppliers is prefetched into the cache. Within 1 hour after the prefetch, only 10% of the prefetched data is accessed, which indicates that the prefetching policy may need to be adjusted.

[0019] Step 202: Determine a candidate cache resource allocation strategy that pairs the first reinforcement learning decision knowledge with the first cache performance statistical data, and an initial correlation weight of the candidate cache resource allocation strategy.

[0020] The first reinforcement learning decision knowledge is obtained by learning the associated cache performance statistical data of the associated online trading platform through the Q-reinforcement learning algorithm, and the associated online trading platform and the target online trading platform belong to the same online trading scenario.

[0021] Next, step 202 involves the step of determining candidate cache resource allocation strategies that pair the first reinforcement learning decision knowledge with the first cache performance statistics data, and initial relevance weights of the candidate cache resource allocation strategies.

[0022] The first reinforcement learning decision knowledge is obtained by learning the associated cache performance statistics of the associated online trading platform through the Q-reinforcement learning algorithm. The associated online trading platform and the target online trading platform in this embodiment belong to the same online trading scenario, for example, both are Internet supply chain financial platforms, but there may be differences in scale, user groups, etc.

[0023] Similar cache performance statistics are also collected on the associated platform, such as data access status information (for example, the frequency of access to supplier credit information on the associated platform is 80 times per hour in a certain period of time), cache resource status data (for example, the total cache capacity of the associated platform is 800MB, 500MB has been used, etc.) and cache policy response information (for example, after the associated platform adopts a certain replacement strategy, the cache hit rate increases from 55% to 58%, etc.).

[0024] Through the Q-reinforcement learning algorithm, a series of decision-making knowledge is learned with the goal of maximizing cache hit rate and maximizing cache resource utilization. For example, it is learned that when the cache fill rate reaches 60% and the access frequency of a certain data suddenly increases, a strategy for prefetching the data should be adopted.

[0025] According to the first cache performance statistics of the target platform, paired with the first reinforcement learning decision knowledge, the candidate cache resource allocation strategy is determined. For example, if the cache fill rate of the target platform is 60%, and the access frequency to the fund flow record suddenly increases in a certain period of time, according to the reinforcement learning decision knowledge, the possible candidate cache resource allocation strategies include giving priority to caching the fund flow record and increasing its storage ratio in the cache.

[0026] For each candidate cache resource allocation strategy, its initial relevance weight also needs to be determined. The initial relevance weight reflects the initial matching degree of the strategy with the current cache performance statistics. For example, there are three candidate cache resource allocation strategies: Strategy A (prioritizes caching of fund flow records), Strategy B (adopts a new replacement strategy), and Strategy C (adjusts the pre-fetching time interval). Based on the current cache performance statistics and reinforcement learning decision knowledge, it is preliminarily estimated that the initial relevance weight of Strategy A is 0.4, the initial relevance weight of Strategy B is 0.3, and the initial relevance weight of Strategy C is 0.3. The determination of this initial weight can be based on some predefined rules, such as a comprehensive evaluation based on factors such as the degree of correlation between data access frequency and strategy, cache resource status and strategy adaptability, etc.

[0027] Step 204: Determine the adjustment characteristics of the candidate cache resource allocation strategy based on the timeliness requirement level of the candidate cache resource allocation strategy.

[0028] Then, step 204 focuses on determining the adjustment characteristics of the candidate cache resource allocation strategy based on the timeliness requirement level of the candidate cache resource allocation strategy.

[0029] It can be understood that each candidate cache resource allocation strategy has its timeliness requirement level. The timeliness requirement level reflects the time sensitivity of the strategy and the urgency of its implementation. For example, for strategy A (prioritize caching of fund flow records), if the access frequency of fund flow records increases sharply in a short period of time, its timeliness requirement level may be high. For example, within 10 minutes, the access frequency of fund flow records increases from 80 times per hour to 150 times per hour. In this case, the cache resource allocation needs to be adjusted as soon as possible to meet the demand.

[0030] For strategy C (adjusting the prefetch interval), if the prefetch strategy has a relatively small impact on the current cache hit rate, its timeliness requirement level may be low. For example, the prefetch operation currently only has a 1%-2% impact on the cache hit rate, and this impact will only be reflected over a long period of time (such as 1-2 days).

[0031] Determine the adjustment characteristics based on the timeliness requirement level. For strategies with high timeliness requirements, the adjustment characteristics may be to quickly increase their relevance weights. For example, for strategy A, since its timeliness requirement level is high, its adjustment characteristics may be to quickly increase the weights based on the current weights at a certain ratio (such as 20%). For strategy C with a low timeliness requirement level, the adjustment characteristics may be to slowly adjust the weights, such as adjusting at a ratio of 5%. For strategy B (adopting a new replacement strategy), if its timeliness requirement level is medium, the adjustment characteristics may be to adjust the weights at a ratio of 10%.

[0032] Step 206: Adjust the initial relevance weight of the candidate cache resource allocation strategy based on the adjustment characteristics of the candidate cache resource allocation strategy to obtain the relevance adjustment weight of the candidate cache resource allocation strategy.

[0033] Furthermore, step 206 can adjust the initial relevance weight of the candidate cache resource allocation strategy based on the adjustment characteristics of the candidate cache resource allocation strategy to obtain the relevance adjustment weight of the candidate cache resource allocation strategy.

[0034] For strategy A, the initial relevance weight is 0.4, and the adjustment feature is to increase the weight by 20%. Then the calculation of its relevance adjustment weight is: (0.4 (1+0.2)=0.48).

[0035] For strategy B, the initial relevance weight is 0.3. Since its adjustment feature is to increase the weight by 10%, its relevance adjustment weight is: (0.3 (1+0.1)=0.33).

[0036] For strategy C, the initial relevance weight is 0.3, and the weight is increased by 5%. The relevance adjustment weight is: (0.3 (1+0.05)=0.315).

[0037] Step 208: Determine a first current cache resource allocation strategy from the candidate cache resource allocation strategies based on the correlation adjustment weights of the candidate cache resource allocation strategies.

[0038] Finally, the first current cache resource allocation strategy may be determined from the candidate cache resource allocation strategies based on the relevance adjustment weights of the candidate cache resource allocation strategies.

[0039] For example, according to the calculated relevance adjustment weight, the candidate cache resource allocation strategy with the highest weight is selected as the first current cache resource allocation strategy. In the above example, the relevance adjustment weight of strategy A is 0.48, that of strategy B is 0.33, and that of strategy C is 0.315. Since strategy A has the highest weight, the first current cache resource allocation strategy is strategy A, that is, the fund flow records are cached first.

[0040] The platform cache performance optimization system will allocate cache resources according to strategy A. For example, it will adjust the storage ratio of fund flow records in the cache and may reallocate part of the cache space originally allocated to other data to fund flow records to improve the cache hit rate and optimize cache performance, thereby improving the operating efficiency of the entire online trading platform in the Internet / supply chain finance scenario.

[0041] Through the above steps, the platform cache performance optimization system can reasonably determine the cache resource allocation strategy based on the actual cache performance statistics of the target online trading platform and the reinforcement learning decision knowledge learned from the associated platform to adapt to the ever-changing online trading environment and improve cache performance.

[0042] It is worth noting that in the online trading platform of Internet / supply chain finance, cache performance optimization is the key to improving the overall efficiency of the system. As a powerful decision-making tool, the Q-reinforcement learning algorithm can learn the optimal cache resource allocation strategy based on the cache performance statistics of the platform. The following will fully and in detail explain the specific application of the Q-reinforcement learning algorithm in each step.

[0043] The technical content of step 200 is to obtain first cache performance statistics of the target online trading platform.

[0044] 1. Data structure and state representation

[0045] State space definition: In the Q-reinforcement learning algorithm, the state space needs to be defined first. For the cache system of the target online trading platform, the state space is constructed by the first cache performance statistics. Data access state information, cache resource state data, and cache policy response information together constitute multiple dimensions that describe the state of the cache system.

[0046] Taking data access status information as an example, the access frequency of each type of data (such as supplier credit information, transaction contract information, capital flow records, etc.) is discretized into several intervals. For example, the access frequency range is 0-10 times / hour for interval 1, 11-20 times / hour for interval 2, etc. In this way, the access frequency of each data type can be represented as a discrete state value. For cache resource status data, such as cache fill rate, it can be divided into intervals such as 0-20%, 21-40%, etc., and cache hit rate can also be discretized in a similar way. Cache policy response information can also be divided into different state levels according to its impact on indicators such as cache hit rate and resource utilization.

[0047] In this way, the state of the entire cache system can be represented as a multi-dimensional vector, for example (S=(s_1, s_2, s_3)), where (s_1) represents the vector corresponding to the data access status information, (s_2) represents the vector corresponding to the cache resource status data, and (s_3) represents the vector corresponding to the cache policy response information.

[0048] Data collection and preprocessing: The platform cache performance optimization system needs to continuously collect these cache performance statistics. For data access status information, the access frequency can be counted at the network layer by monitoring the target object of the data request (that is, the specific data type) and the timestamp of the request. For example, a counter is used to accumulate the number of accesses for each data type, and the access frequency data is updated every certain period of time (such as 1 hour).

[0049] Cache resource status data can be directly obtained from the cache management module, including the total capacity of the cache, the used capacity, and the space occupied by different types of data in the cache. Cache policy response information needs to be obtained after each cache policy execution (such as cache replacement and pre-fetch operations) by comparing the changes in indicators such as cache hit rate and resource utilization before and after the operation.

[0050] After collecting the raw data, it needs to be preprocessed to meet the requirements of the Q-reinforcement learning algorithm. As mentioned above, the discretization process converts the previous and subsequent related values ​​into discrete state values. This helps to reduce the size of the state space and improve the learning efficiency of the algorithm.

[0051] 2. Data Utilization in Q-Reinforcement Learning Algorithm

[0052] As learning input: These processed first cache performance statistics are used as input to the Q-reinforcement learning algorithm, providing the algorithm with a comprehensive description of the current cache system status. The algorithm evaluates the pros and cons of different cache resource allocation strategies based on this status information.

[0053] For example, when the algorithm receives a status indicating that the current cache fill rate is high (such as in the 60-80% range), the access frequency of certain key data (such as fund flow records) has suddenly increased (entered a higher access frequency range), and the current cache strategy has caused the cache hit rate to drop slightly (at a lower cache strategy response level), it will look for a cache resource allocation strategy that can improve this situation in subsequent steps.

[0054] Step 202 involves determining a candidate cache resource allocation strategy that pairs the first reinforcement learning decision knowledge with the first cache performance statistical data, and a related technical solution for initial relevance weights of the candidate cache resource allocation strategy.

[0055] 1.1 Generation of Decision Knowledge in First Reinforcement Learning

[0056] Learning process of the associated platform: The first reinforcement learning decision knowledge is obtained by learning the associated cache performance statistics of the associated online trading platform through the Q-reinforcement learning algorithm. On the associated platform, a similar state space is also constructed to represent the state of the cache system.

[0057] For example, the learning process of the Q-reinforcement learning algorithm of the association platform is as follows: the algorithm initializes a Q-table, where the rows represent different states and the columns represent different cache resource allocation strategies (such as caching certain data first, adopting a specific cache replacement strategy, adjusting the prefetch strategy, etc.). During the learning process, the algorithm updates the Q-value in the Q-table based on the reward feedback after each execution of the cache resource allocation strategy.

[0058] For example, when the associated platform is in a state indicating a medium cache fill rate (e.g., 40-60%) and a high frequency of access to transaction contract information (in a high access frequency interval), the algorithm tries a strategy of caching transaction contract information first. If the cache hit rate is improved (a positive reward, such as +1) after executing the strategy, the algorithm will update the corresponding state-strategy pair in the Q-table (i.e., the Q-value of executing the strategy of caching transaction contract information first in the current state) according to the Q-learning formula (Q(s,a)=Q(s,a)+α[r+γmax_{a'}Q(s',a')-Q(s,a)]) (where (α) is the learning rate, (r) is the reward, (γ) is the discount factor, (s) is the current state, (a) is the executed strategy, (s') is the next state, and (a') is all possible strategies in the next state).

[0059] Through a large amount of state-strategy interactive learning on the associated platform, the Q-reinforcement learning algorithm gradually converges and forms the first reinforcement learning decision knowledge, that is, the more optimal cache resource allocation strategy under different states and its corresponding Q-value.

[0060] Pairing and initial correlation weight determination: When faced with the first cache performance statistics of the target online trading platform, the platform cache performance optimization system pairs the state represented by these data with the first reinforcement learning decision knowledge. For example, if the state of the target platform is similar to a state in the associated platform (with a high degree of similarity in dimensions such as data access state information, cache resource state data, and cache policy response information), then the cache resource allocation strategy associated with the state is determined as a candidate cache resource allocation strategy.

[0061] The initial relevance weight of the candidate cache resource allocation strategy is determined based on the Q-value learned by the Q-reinforcement learning algorithm on the associated platform. For example, in the Q-table learned on the associated platform, the Q-value of a candidate strategy in a similar state is 0.6, and the sum of the Q-values ​​of all candidate strategies in this state is 1.5, then the initial relevance weight of the candidate strategy can be calculated as (0.6 / 1.5=0.4). This calculation method based on the Q-value ratio reflects the relative pros and cons of the strategy in the learning process of the associated platform, and serves as a reference for the initial relevance weight on the target platform.

[0062] Step 204 includes the steps of determining the adjustment characteristics of the candidate cache resource allocation strategy based on the timeliness requirement level of the candidate cache resource allocation strategy.

[0063] 1. The relationship between quantification of time-sensitive requirements and Q-reinforcement learning

[0064] Quantitative representation of timeliness requirements: In the framework of the Q-reinforcement learning algorithm, timeliness requirements can also be quantified. For example, timeliness requirements can be divided into three levels: high, medium, and low, represented by the values ​​3, 2, and 1, respectively.

[0065] The timeliness requirement level is determined based on dynamic analysis of cache performance statistics. For data access status information, if the growth rate of access frequency of a certain data exceeds a certain threshold (such as 50%) within a short period of time (such as within 10 minutes), then the timeliness requirement level of the cache resource allocation strategy related to the data (such as caching the data first) may be determined to be high. For cache resource status data, if the cache fill rate is close to the upper limit (such as more than 80%), and the current cache strategy causes the cache hit rate to drop rapidly, then the timeliness requirement level for adjusting the cache strategy (such as adopting a new replacement strategy) may be high.

[0066] This quantified time requirement level is associated with the reward mechanism in the Q-RL algorithm. For example, for a strategy with a high time requirement level, if timely execution can effectively improve cache performance (such as increasing cache hit rate, reducing resource waste, etc.), then executing the strategy will receive a higher reward.

[0067] Determine the adjustment characteristics: Determining the adjustment characteristics according to the time requirement level is a strategy adjustment mechanism of the Q-reinforcement learning algorithm in dealing with dynamic environments. For candidate cache resource allocation strategies with a high time requirement level (value 3), the adjustment characteristics may be to adjust the relevance weights to a large extent, such as increasing the weights by 20%. This can be achieved by introducing coefficients related to the time requirement level in the Q-learning formula. For example, when updating the Q-value, for strategies with a high time requirement level, (α) (learning rate) is multiplied by a coefficient greater than 1 (such as 1.2) to accelerate the adjustment of weights.

[0068] For a strategy with a medium time requirement (value 2), the adjustment feature might be to adjust the weight by 10%, which corresponds to (α) multiplied by 1.1 in the Q-learning formula. For a strategy with a low time requirement (value 1), the adjustment feature might be to adjust the weight by 5%, which corresponds to (α) multiplied by 1.05. This method of determining the adjustment feature based on the time requirement level enables the Q-reinforcement learning algorithm to adjust the weights of candidate cache resource allocation strategies in a targeted manner based on the real-time needs of the cache system.

[0069] Step 206 involves adjusting the initial relevance weight of the candidate cache resource allocation strategy based on the adjustment characteristics of the candidate cache resource allocation strategy to obtain the relevance adjustment weight of the candidate cache resource allocation strategy.

[0070] 1. Weight Adjustment Calculation and Q-Learning Update

[0071] Specific calculation of weight adjustment: The initial relevance weight of the candidate cache resource allocation strategy is adjusted based on the adjustment feature determined in step 204. For example, there is a candidate cache resource allocation strategy A, whose initial relevance weight (w_0=0.4), the timeliness requirement level of the strategy is high, and the adjustment feature is to increase the weight by 20%.

[0072] Then the calculation of its correlation adjustment weight (w_1) is: (w_1=w_0 (1+0.2)=0.4 1.2=0.48). This calculation process can be regarded as the update of the strategy weight according to the environmental feedback (i.e., the adjustment characteristics corresponding to the immediate demand level) in the Q-reinforcement learning algorithm.

[0073] If another candidate strategy B has an initial relevance weight (w_0'=0.3), a timeliness requirement level of medium, and an adjustment feature of increasing the weight by 10%, then its relevance adjustment weight (w_1'=w_0' (1+0.1)=0.3 1.1=0.33).

[0074] Integration with Q-learning algorithm: In the Q-reinforcement learning algorithm, this weight adjustment process is continuously iterated. Each time, the timeliness requirement level is re-evaluated based on cache performance statistics (including data access status information, cache resource status data, and cache policy response information updates), and the relevance weight is adjusted accordingly.

[0075] For example, as time goes by, if the data access pattern changes, a strategy that originally had a medium time requirement level may become a high time requirement level. In this case, the next weight adjustment will be made according to the adjustment characteristics of the high time requirement level. This dynamic weight adjustment mechanism enables the Q-reinforcement learning algorithm to continuously adapt to the dynamic changes of the cache system and gradually converge to the optimal cache resource allocation strategy.

[0076] Step 208 involves a technical solution of determining a first current cache resource allocation strategy from the candidate cache resource allocation strategies based on the correlation adjustment weights of the candidate cache resource allocation strategies.

[0077] 1. Strategy selection and Q-learning goal achievement

[0078] Strategy selection basis: Under the framework of the Q-reinforcement learning algorithm, the weights are adjusted according to the calculated relevance of the candidate cache resource allocation strategies, and the strategy with the highest weight is selected as the first current cache resource allocation strategy. This selection process is based on the goal of the Q-reinforcement learning algorithm to maximize the long-term cumulative reward.

[0079] For example, there are three candidate cache resource allocation strategies A, B, and C, whose relevance adjustment weights are (0.48), (0.33), and (0.315) respectively. Since strategy A has the highest weight, strategy A is selected as the first current cache resource allocation strategy.

[0080] From the perspective of Q-learning, choosing the strategy with the highest weight means that in the current state, the strategy is considered to be the strategy most likely to obtain the highest long-term cumulative reward. This reward is reflected in the improvement of cache performance, such as increasing cache hit rate and optimizing cache resource utilization.

[0081] Impact on the cache system and continuous learning: Once the first current cache resource allocation strategy is selected, the platform cache performance optimization system will allocate cache resources according to the strategy. For example, if strategy A is to cache fund flow records first, then the system will adjust the storage ratio of fund flow records in the cache, and may allocate more cache space to fund flow records.

[0082] At the same time, the Q-RL algorithm will continue to monitor changes in cache performance statistics. After executing strategy A, the state of the cache system will change, and the new cache performance statistics will serve as the input for the next round of Q-RL algorithm. Based on the new state, the algorithm will re-evaluate the relevance weights of different strategies and continue to optimize the cache resource allocation strategy to adapt to the ever-changing online transaction environment.

[0083] Through the above detailed description, it can be seen that the Q-reinforcement learning algorithm plays a key decision-making and optimization role in steps 200 to 208, and can effectively improve the cache performance of the Internet / supply chain finance online transaction platform.

[0084] In summary, the embodiment of the present invention optimizes the cache resource allocation strategy of the online trading platform through a series of steps. First, comprehensive cache performance statistics are obtained to provide a basis for accurate decision-making. The candidate strategies and initial weights are determined using the reinforcement learning decision knowledge learned from the associated platform, and the weights are adjusted with the help of the timeliness requirement level to finally determine the optimal strategy. This process can dynamically adapt to the platform's data access mode and cache status changes, improve the cache hit rate, optimize cache resource utilization, reduce system response delays caused by improper cache strategies, and enhance the overall performance and user experience of the online trading platform in the Internet / supply chain finance scenario.

[0085] In a preferred embodiment, step 204 involves determining the adjustment characteristics of the candidate cache resource allocation policy based on the timeliness requirement level of the candidate cache resource allocation policy, including: determining the timeliness requirement level as the current cache performance optimization decision weight of the candidate cache resource allocation policy; obtaining the past policy execution trust weight and the past cache performance optimization decision weight of the candidate cache resource allocation policy; determining the current policy execution trust weight of the candidate cache resource allocation policy based on the timeliness requirement level and the past policy execution trust weight; determining the current platform mode matching weight of the candidate cache resource allocation policy based on the timeliness requirement level and the past cache performance optimization decision weight; determining the adjustment characteristics of the candidate cache resource allocation policy based on the current cache performance optimization decision weight, the current policy execution trust weight and the current platform mode matching weight of the candidate cache resource allocation policy.

[0086] In some optional embodiments, before determining the timeliness requirement level as the current cache performance optimization decision weight of the candidate cache resource allocation strategy, it also includes: obtaining a setting association mechanism between an online transaction session cycle and a cache timeliness tolerance; based on the setting association mechanism, determining a target cache timeliness tolerance corresponding to the current online transaction session cycle; obtaining a cache timeliness warning score of the candidate cache resource allocation strategy; and determining the timeliness requirement level of the candidate cache resource allocation strategy based on the target cache timeliness tolerance and the cache timeliness warning score. Based on this, the obtaining of the past policy execution trust weight and the past cache performance optimization decision weight of the candidate cache resource allocation policy includes: determining the past target cache time tolerance corresponding to the past online transaction session cycle based on the set association mechanism; obtaining the past cache time warning score of the candidate cache resource allocation policy; determining the past time requirement level of the candidate cache resource allocation policy based on the past target cache time tolerance and the past cache time warning score; determining the past policy execution trust weight and the past cache performance optimization decision weight of the candidate cache resource allocation policy based on the past time requirement level.

[0087] In the embodiments of the present invention, the above preferred embodiments are described in detail as follows.

[0088] 1. Timeliness requirement level and current cache performance optimization decision weight

[0089] Preliminary determination of timeliness requirement level: In this embodiment, the timeliness requirement level is first determined. In order to determine the timeliness requirement level of the candidate cache resource allocation strategy, it is necessary to obtain the setting association mechanism of the online transaction session cycle and the cache timeliness tolerance.

[0090] For example, the online trading platform sets up such an association mechanism: when the online trading session cycle is 0-10 minutes, the cache time tolerance is 5 minutes; when the online trading session cycle is 10-30 minutes, the cache time tolerance is 10 minutes; when the online trading session cycle is 30-60 minutes, the cache time tolerance is 15 minutes, etc. This is a setting based on different transaction durations with different requirements for cache data timeliness.

[0091] For example, the current online transaction session cycle is 20 minutes. According to the above-mentioned association mechanism, the corresponding target cache time tolerance is 10 minutes. Then obtain the cache time warning score of the candidate cache resource allocation strategy. For example, for a candidate cache resource allocation strategy (such as caching funds flow records first), its cache time warning score is obtained by comprehensively evaluating the timeliness-related indicators of this type of data in the cache (such as data update frequency, the interval between the last update time and the current time, etc.). For example, if the data update frequency of the funds flow record in the cache is high, and 8 minutes have passed since the last update (close to the cache time tolerance of 10 minutes), its cache time warning score may be 0.8 (a value between 0-1, the closer to 1, the closer to the time tolerance limit).

[0092] The timeliness requirement level is determined based on the target cache timeliness tolerance and cache timeliness warning score. An exemplary formula can be set: Timeliness requirement level = cache timeliness warning score (10 / target cache time tolerance). In the above example, time requirement level = 0.8 10 / 10=0.8 (here, for example, the timeliness requirement level is also expressed as a value between 0 and 1).

[0093] Then the timeliness requirement level is determined as the current cache performance optimization decision weight of the candidate cache resource allocation strategy. This means that the higher the timeliness requirement level, the greater the weight of the strategy in the current cache performance optimization decision, reflecting its importance to the current cache performance improvement.

[0094] Determination of past timeliness requirement level and weight: When determining the past policy execution trust weight and the past cache performance optimization decision weight, it is also based on the set association mechanism. First, determine the past target cache timeliness tolerance corresponding to the past online transaction session cycle. For example, in a previous 15-minute online transaction session cycle, the corresponding past target cache timeliness tolerance is 10 minutes.

[0095] Get the past cache timeliness warning score of the candidate cache resource allocation strategy. For example, for the previously mentioned priority cache fund flow record strategy, the past cache timeliness warning score is 0.6. According to the aforementioned timeliness requirement level calculation formula, the past timeliness requirement level = 0.6 (10 / 10)=0.6.

[0096] Based on the past timeliness requirement level, the past policy execution trust weight and the past cache performance optimization decision weight are determined. For example, there is such a relationship: Past policy execution trust weight = past timeliness requirement level 0.5, past cache performance optimization decision weight = past timeliness requirement level 0.3. So for this example, the past policy execution trust weight = 0.6 0.5=0.3, past cache performance optimization decision weight = 0.6 0.3=0.18.

[0097] 2. Determination of the current policy execution trust weight

[0098] Based on the timeliness requirement level and past policy execution trust weight: The current policy execution trust weight of the candidate cache resource allocation policy is determined based on the timeliness requirement level and past policy execution trust weight. A weighted average formula can be used here:

[0099] Current policy execution trust weight = (timeliness requirement level 0.6+ past strategy execution trust weight 0.4).

[0100] In the previous example, the timeliness requirement level is 0.8, and the past policy execution trust weight is 0.3. Then the current policy execution trust weight = 0.8 0.6+0.3 0.4=0.48+0.12=0.6.

[0101] The 0.6 and 0.4 in this formula are weight coefficients set based on the relative importance of the current timeliness requirement and the past execution trust. 0.6 means that the current timeliness requirement level has a higher relative importance in determining the current strategy execution trust weight, while 0.4 means that the past strategy execution trust weight also has a certain influence.

[0102] 3. Determination of the current platform model matching weight

[0103] Optimize decision weights based on timeliness requirement level and past cache performance: The current platform pattern matching weights for determining candidate cache resource allocation strategies are based on timeliness requirement level and past cache performance optimization decision weights. The formula can be used:

[0104] Current platform mode matching weight = timeliness requirement level 0.7+ Past cache performance optimization decision-making power 0.3.

[0105] The timeliness requirement level has been calculated to be 0.8, and the past cache performance optimization decision weight is 0.18. So the current platform mode matching weight = 0.8 0.7+0.18 0.3=0.56+0.054=0.614.

[0106] The 0.7 and 0.3 here are coefficients set based on the relative importance of the current timeliness requirement and the past cache performance optimization decision weights in matching the platform mode. 0.7 indicates that the current timeliness requirement level has a greater impact on the current platform mode matching weight, and 0.3 indicates that the past cache performance optimization decision weights also play a certain role.

[0107] 4. Determination of regulatory characteristics

[0108] Determine adjustment characteristics based on multiple weights: Finally, determine the adjustment characteristics of the candidate cache resource allocation strategy based on the current cache performance optimization decision weight, the current policy execution trust weight and the current platform pattern matching weight of the candidate cache resource allocation strategy.

[0109] For example, the adjustment feature is represented by a comprehensive weight, where the comprehensive weight = the current cache performance optimization decision weight 0.3+Current policy execution trust weight 0.3+Current platform pattern matching weight 0.4.

[0110] In the previous example, the current cache performance optimization decision weight is 0.8, the current policy execution trust weight is 0.6, and the current platform pattern matching weight is 0.614. The combined weight = 0.8 0.3+0.6 0.3+0.614 0.4=0.24+0.18+0.2456=0.6656.

[0111] This comprehensive weight represents the adjustment feature, and the initial relevance weight of the candidate cache resource allocation strategy can be adjusted based on this comprehensive weight. For example, if the initial relevance weight is w_0, then the adjusted weight w_1=w_0 (1+0.6656) (Here, for example, the weight is increased proportionally, and the specific adjustment method can be set according to actual needs).

[0112] In this way, the adjustment characteristics of the candidate cache resource allocation strategy are determined through a series of complex but effective steps. First, when determining the timeliness requirement level, the relationship between the online transaction session cycle and the cache timeliness tolerance and the cache timeliness warning score are fully considered, so that the timeliness requirement level can accurately reflect the degree of timeliness requirement of the current cache resource allocation strategy. Then, when determining the current strategy execution trust weight and the current platform pattern matching weight, the timeliness requirement level and the past relevant weight are combined, considering both the importance of the current situation and past experience. The adjustment characteristics finally determined by these weights can more accurately adjust the initial relevance weight of the candidate cache resource allocation strategy. This helps to improve the adaptability and accuracy of the cache resource allocation strategy and optimize cache performance. For example, the weight adjustment determined in this way can improve the cache hit rate, reduce the system performance degradation caused by untimely or unreasonable allocation of cache data, and improve the operation efficiency of the entire online trading platform.

[0113] In another optional technical solution, after the first current cache resource allocation strategy is determined from the candidate cache resource allocation strategies based on the correlation adjustment weights of the candidate cache resource allocation strategies, it also includes: obtaining second cache performance statistical data of the target online trading platform; determining a target cache performance cluster feature set corresponding to the second cache performance statistical data based on a set association mechanism between cache performance statistical data and cache performance cluster feature sets, wherein the cache performance cluster feature set is obtained by performing feature clustering on the cache performance statistical data corresponding to it; determining a target learning reward quantitative feature corresponding to the target cache performance cluster feature set based on a set association mechanism between cache performance cluster feature sets and learning reward quantitative features; determining the resource allocation optimization strategy corresponding to the target learning reward quantitative feature as the second current cache resource allocation strategy; determining a third current cache resource allocation strategy from the first current cache resource allocation strategy and the second current cache resource allocation strategy; and recommending the third current cache resource allocation strategy to the target online trading platform.

[0114] In addition, before obtaining the second cache performance statistical data of the target online trading platform, it also includes: obtaining the second cache performance statistical data of several online trading platforms to be processed, and obtaining several cache performance statistical data to be processed; performing feature clustering on several of the cache performance statistical data to be processed, and obtaining several cache performance cluster feature sets; determining the cache performance statistical data to be processed of each of the cache performance cluster feature sets obtained by feature clustering as the cache performance statistical data to be processed corresponding to each of the cache performance cluster feature sets, and obtaining a set association mechanism between the cache performance statistical data and the cache performance cluster feature sets.

[0115] 1. Obtaining the Second Cache Performance Statistics and Determining the Target Cache Performance Cluster Feature Set

[0116] Acquisition of second cache performance statistics: After determining the first current cache resource allocation strategy, it is necessary to obtain the second cache performance statistics of the target online trading platform. This data is similar to the first cache performance statistics in the previous step, and also includes aspects such as data access status information, cache resource status data, and cache strategy response information. For example, in the Internet / supply chain finance scenario, data access status information may include the frequency of access to data related to different supply chain links (such as order information, logistics information, payment information, etc.). For example, within a period of time, the access frequency of order information is 150 times per hour, logistics information is 80 times per hour, and payment information is 100 times per hour; cache resource status data such as the total cache capacity of 1000MB, the used capacity of 600MB, the cache hit rate of 60%, etc.; cache strategy response information such as the change in cache hit rate after adopting a certain cache replacement strategy, etc.

[0117] Determination of cache performance cluster feature set: Determine the target cache performance cluster feature set corresponding to the second cache performance statistical data based on the set association mechanism between the cache performance statistical data and the cache performance cluster feature set. Before obtaining this set association mechanism, it is necessary to first perform feature clustering on the second cache performance statistical data of several to-be-processed online trading platforms (i.e., several to-be-processed cache performance statistical data).

[0118] For example, a feature clustering method based on a clustering algorithm (such as the K-means algorithm) is used. For each pending cache performance statistical data, it is regarded as a data point. The entity quantitative knowledge of the data point consists of dimensions such as data access status information, cache resource status data, and cache strategy response information. For example, for a pending cache performance statistical data, its data access status information vector is ((150, 80, 100)) (representing the access frequency of order, logistics, and payment information, respectively), the cache resource status data vector is ((1000, 600, 0.6)) (representing the total capacity, used capacity, and hit rate, respectively), and the cache strategy response information vector is ((-0.05)) (representing the change in hit rate after adopting a certain strategy).

[0119] Use the K-means algorithm and set the number of cluster centers (K=3) (the value of (K) here is determined based on actual conditions and experience, with the goal of dividing cache performance statistics with similar characteristics into (K) clusters). The goal of the algorithm is to minimize the sum of the distances from each data point to its cluster center. The distance metric can be the Euclidean distance formula (d=sqrt{sum_{i=1}^{n}(x_{i}-y_{i})^{2}}), where (x_{i}) and (y_{i}) are the values ​​of the two data points in the (i)th dimension, and (n) is the number of dimensions of the data points.

[0120] After feature clustering, several cache performance cluster feature sets are obtained. For example, one cache performance cluster feature set may contain cache performance statistics with high cache hit rates and relatively stable data access frequencies; another cache performance cluster feature set may contain statistics with low cache hit rates and large fluctuations in data access frequencies.

[0121] Then, according to the set association mechanism between cache performance statistics and cache performance cluster feature sets, the target cache performance cluster feature set corresponding to the second cache performance statistics of the target online trading platform is determined. This association mechanism can be a mapping relationship. For example, if the distance between the second cache performance statistics of the target online trading platform and the central entity quantitative knowledge of a cache performance cluster feature set (obtained by calculating the average entity quantitative knowledge of all data points in the cluster) in terms of key indicators such as data access frequency and cache hit rate is less than a certain threshold (such as 0.5, which is also determined based on actual conditions), it is determined as the corresponding target cache performance cluster feature set.

[0122] 2. Target learning reward quantization characteristics and determination of the second current cache resource allocation strategy

[0123] Determination of target learning reward quantitative features: Based on the association mechanism between cache performance cluster feature set and learning reward quantitative features, determine the target learning reward quantitative features corresponding to the target cache performance cluster feature set. For example, the learning reward quantitative features are set according to the optimization goals of cache performance, such as improving cache hit rate, reducing cache resource waste, etc.

[0124] For different cache performance cluster feature sets, different learning reward quantization features are set. For example, for a cache performance cluster feature set with a high and stable cache hit rate, its corresponding learning reward quantization feature may focus on further optimizing cache resource allocation to reduce resource waste. A learning reward quantization feature formula can be set: (R=0.6 (1-frac{used cache capacity}{total cache capacity})+0.4 cache hit rate) (where (R) is the learning reward quantization feature, and the parameters (0.6) and (0.4) in this formula are weights determined according to the relative importance of cache resource utilization and cache hit rate in reward quantification).

[0125] If the target cache performance cluster feature set belongs to the type with high and stable cache hit rate, the corresponding target learning reward quantitative feature is calculated according to the above formula. For example, if the total cache capacity is 1000MB, the used cache capacity is 400MB, and the cache hit rate is 80%, then (R=0.6 (1-frac{400}{1000})+0.4 0.8=0.6 0.6+0.32=0.68).

[0126] Determination of the second current cache resource allocation strategy: The resource allocation optimization strategy corresponding to the target learning reward quantitative feature is determined as the second current cache resource allocation strategy. This means selecting a resource allocation optimization strategy that best suits the current cache performance cluster feature set based on the learning reward quantitative feature. For example, if the learning reward quantitative feature (R) is high, the corresponding resource allocation optimization strategy may be to prioritize allocating more cache resources to data types that contribute more to the cache hit rate improvement (such as key data such as order information and payment information in the Internet / supply chain finance).

[0127] 3.3 Determination and recommendation of current cache resource allocation strategy

[0128] Determination of the third current cache resource allocation strategy: Determine the third current cache resource allocation strategy from the first current cache resource allocation strategy and the second current cache resource allocation strategy. A weighted selection method can be used. For example, the weight of the first current cache resource allocation strategy is (w_1=0.4) (this weight can be determined based on the reliability, timeliness and other factors of the first current cache resource allocation strategy), and the weight of the second current cache resource allocation strategy is (w_2=0.6) (determined in the same way).

[0129] For example, the first current cache resource allocation strategy is to prioritize caching fund flow records, and the second current cache resource allocation strategy is to prioritize allocating more cache resources to order information and payment information. A comprehensive resource allocation strategy can be calculated based on weights. For example, the decision on cache resource allocation is measured according to the proportion of data types in the cache. For fund flow records, the proportion under the first current cache resource allocation strategy is (p_1=0.3), and for order information and payment information, the proportions under the second current cache resource allocation strategy are (p_2=0.4) and (p_3=0.3), respectively.

[0130] Then under the third current cache resource allocation strategy,

[0131] The proportion of fund flow records in the cache (p_{1}'=p_1 w_1=0.3 0.4=0.12),

[0132] The proportion of order information (p_{2}'=p_2 w_2=0.4 0.6=0.24),

[0133] The proportion of payment information (p_{3}'=p_3 w_2=0.3 0.6=0.18).

[0134] Recommending the third current cache resource allocation strategy: Finally, recommending the third current cache resource allocation strategy to the target online trading platform. The platform adjusts cache resource allocation according to the recommended strategy to improve cache performance.

[0135] It can be seen that by obtaining the second cache performance statistics and determining its corresponding target cache performance cluster feature set, the cache performance status can be analyzed in more detail. Determining the second current cache resource allocation strategy by using the association mechanism between the cache performance cluster feature set and the learning reward quantification feature helps to formulate more targeted strategies according to different cache performance conditions. Determining the third current cache resource allocation strategy from the first and second current cache resource allocation strategies and recommending it to the target online trading platform combines the advantages of different strategies. For example, in the Internet / supply chain finance scenario, it is possible to reasonably allocate cache resources according to the actual data access situation (such as the access frequency and importance of information such as orders, logistics, and payments), improve the cache hit rate, optimize cache resource utilization, and reduce the system performance degradation caused by unreasonable cache strategies, thereby improving the operating efficiency of the entire online trading platform.

[0136] In a replaceable embodiment, after performing feature clustering on a number of cache performance statistical data to be processed to obtain a number of cache performance cluster feature sets, it also includes: obtaining a number of reinforcement learning decision knowledge to be processed; binding a number of reinforcement learning decision knowledge to be processed with a number of cache performance cluster feature sets one by one to obtain a number of knowledge feature binary groups, each of which includes a reinforcement learning decision knowledge to be processed and a cache performance cluster feature set; determining the commonality measurement coefficient between the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set in each of the knowledge feature binary groups to obtain the commonality measurement coefficient corresponding to each of the knowledge feature binary groups; walking the cache performance cluster feature sets in the knowledge feature binary groups corresponding to each of the commonality measurement coefficients in descending order to determine the cache performance cluster feature set corresponding to each of the reinforcement learning decision knowledge to be processed, and obtaining a set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge.

[0137] 1. Binding of the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set

[0138] Acquisition of pending reinforcement learning decision knowledge: After performing feature clustering on several pending cache performance statistics to obtain several cache performance cluster feature sets, it is necessary to first obtain several pending reinforcement learning decision knowledge. These pending reinforcement learning decision knowledge are obtained through the Q-reinforcement learning algorithm in different learning processes. For example, in the online trading platform scenario of Internet / supply chain finance, these decision knowledge may be about cache resource allocation strategies for different data types (such as supplier credit data, transaction order data, logistics information data, etc.) under different cache states (such as different cache fill rates, different data access frequency distributions, etc.).

[0139] For example, there are three reinforcement learning decision knowledge to be processed: decision knowledge A, decision knowledge B, and decision knowledge C. Decision knowledge A may be a strategy of giving priority to caching supplier credit data when the cache filling rate is low (such as 30%) and the access frequency to supplier credit data is high; decision knowledge B may be a strategy of caching these two data in a certain ratio (such as 1:1) when the cache filling rate is medium (such as 50%) and the access frequency of transaction order data and logistics information data is similar; decision knowledge C may be a strategy of giving priority to caching logistics information data and adopting a specific cache replacement strategy when the cache filling rate is high (such as 70%) and the access frequency of logistics information data is extremely high.

[0140] Binding and formation of knowledge feature binary groups: Bind several pieces of reinforcement learning decision knowledge to be processed with several cache performance cluster feature sets one by one to obtain several knowledge feature binary groups. For example, after feature clustering, three cache performance cluster feature sets are obtained: feature set 1, feature set 2, and feature set 3.

[0141] Then the knowledge feature tuples formed may include: (decision knowledge A, feature set 1), (decision knowledge B, feature set 2), (decision knowledge C, feature set 3), etc. Each knowledge feature tuple contains a to-be-processed reinforcement learning decision knowledge and a cache performance cluster feature set. This binding lays the foundation for the subsequent determination of the relationship between them.

[0142] 2. Determination of commonality measurement coefficient

[0143] Calculation method of commonality metric coefficient: determine the commonality metric coefficient of the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set in each knowledge feature binary. In order to calculate this commonality metric coefficient, a method based on entity quantified knowledge similarity can be used.

[0144] First, both the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set are converted into entity quantitative knowledge. For the reinforcement learning decision knowledge to be processed, its entity quantitative knowledge can be composed of cache-related factors (such as cache state factors, data type factors, etc.). For example, the entity quantitative knowledge corresponding to decision knowledge A (V_A=(0.3, 1, 0, 0)) (here 0.3 represents the cache filling rate, 1 represents the importance weight of supplier credit data, 0 represents the weight of other data types, and 0 represents a specific cache strategy identifier).

[0145] For the cache performance cluster feature set, its entity quantitative knowledge can be composed of the characteristics of cache performance statistical data within the cluster (such as average cache fill rate, access frequency ratio of different data types, etc.). For example, the entity quantitative knowledge corresponding to feature set 1 (V_1=(0.25, 0.8, 0.1, 0)) (here 0.25 represents the average cache fill rate, 0.8 represents the access frequency ratio of supplier credit data, 0.1 represents the access frequency ratio of other data types, and 0 represents specific cache strategy-related features).

[0146] Then, the cosine similarity formula can be used to calculate the commonality measure coefficient. The cosine similarity formula is:

[0147] (sim(V_A, V_1)

[0148] =frac{V_AcdotV_1}{vertV_AvertvertV_1vert});

[0149] Where (V_AcdotV_1) is the vector dot product, (vertV_Avert) and (vertV_1vert) are the magnitudes of the vectors.

[0150] The calculation can be obtained:

[0151] (sim(V_A, V_1)

[0152] =frac{0.3 0.25+1 0.8+0 0.1+0 0}{sqrt{0.3^2+1^2+0^2+0^2} sqrt{0.25^2+0.8^2+0.1^2+0^2}}).

[0153] For example, after calculation, we get (sim(V_A, V_1)=0.8), which is the commonality measurement coefficient corresponding to the knowledge feature binary (decision knowledge A, feature set 1). The same method can be used to calculate the commonality measurement coefficients corresponding to other knowledge feature binary pairs.

[0154] 3. Determination of the wandering and association mechanism of cache performance cluster feature sets

[0155] According to the walking process of the commonality measurement coefficient: Walk the cache performance cluster feature set in each knowledge feature binary corresponding to the commonality measurement coefficient in descending order. For example, the commonality measurement coefficient of (decision knowledge A, feature set 1) is calculated to be 0.8, the commonality measurement coefficient of (decision knowledge B, feature set 2) is 0.6, and the commonality measurement coefficient of (decision knowledge C, feature set 3) is 0.7.

[0156] Arrange in descending order: (decision knowledge A, feature set 1), (decision knowledge C, feature set 3), (decision knowledge B, feature set 2). Then start from the highest commonality metric coefficient. For decision knowledge A, since its commonality metric coefficient with feature set 1 is the highest, it is preliminarily determined that feature set 1 is associated with decision knowledge A.

[0157] During the walking process, if a decision knowledge has a high commonality measurement coefficient with multiple feature sets (for example, there is another feature set 4 whose commonality measurement coefficient with decision knowledge A is also 0.75), further judgment can be made based on other factors. For example, the timeliness factor of the data in the feature set can be considered. If decision knowledge A pays more attention to the timeliness of the data, and the timeliness of the data in feature set 1 is more in line with the requirements of decision knowledge A (for example, by comparing the update frequency of the data in feature set 1 and feature set 4, etc.), then feature set 1 is still determined to be associated with decision knowledge A.

[0158] The association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge: Through such a walking process, the cache performance cluster feature set corresponding to each reinforcement learning decision knowledge to be processed is determined, and the association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge is obtained. This association mechanism is a mapping relationship, which clarifies which reinforcement learning decision knowledge should be used under different cache performance cluster feature sets. For example, when the cache performance cluster feature set is feature set 1, decision knowledge A should be used; when it is feature set 2, decision knowledge B should be used, and so on. This association mechanism helps to more accurately select appropriate reinforcement learning decision knowledge according to cache performance conditions in the subsequent cache resource allocation strategy determination process, thereby optimizing the cache resource allocation strategy.

[0159] Therefore, by determining the commonality metric coefficient and determining the association mechanism according to its descending walk, the reinforcement learning decision knowledge can be more accurately matched with the cache performance cluster feature set. For example, in the Internet / supply chain finance scenario, accurate matching helps to select the most suitable cache resource allocation strategy for different cache states (such as cache fill rate, data access frequency, etc.) and data types (such as supplier credit data, transaction order data, etc.). This can improve the cache hit rate, because the appropriate decision knowledge can guide the cache resources to be more reasonably allocated to high-demand data. At the same time, it optimizes the utilization efficiency of cache resources, reduces the waste of cache resources and system performance degradation caused by inappropriate strategy selection, and thus improves the operating efficiency of the entire online trading platform.

[0160] Under a preferred design idea, the cache performance cluster feature sets in the knowledge feature binary corresponding to each commonality measurement coefficient are wandered in descending order of the commonality measurement coefficient to determine the cache performance cluster feature set corresponding to each of the reinforcement learning decision knowledge to be processed, and a set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge is obtained, including: wandering the cache performance cluster feature sets in the learning reward quantization feature pairs corresponding to each of the commonality measurement coefficients in descending order of the commonality measurement coefficient, and judging whether the number of reinforcement learning decision knowledge to be processed corresponding to the wandered target cache performance cluster feature set reaches the set number of learning reward quantization features, and whether the number of cache performance cluster feature sets corresponding to the target reinforcement learning decision knowledge to be processed in the knowledge feature binary where the wandered target cache performance cluster feature set is located reaches the set number of cache performance cluster feature sets; if so, the target cache performance cluster feature set is determined as the cache performance cluster feature set corresponding to the target reinforcement learning decision knowledge to be processed, and wandering continues to the next knowledge The cache performance cluster feature set in the feature binary group is selected, and the step of jumping to judging whether the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set of the wandering reaches the set number of quantitative features of learning rewards, and whether the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge in the knowledge feature binary group where the target cache performance cluster feature set of the wandering is located reaches the set number of cache performance cluster feature sets; if not, the cache performance cluster feature set in the next knowledge feature binary group is continued to wander, and the step of jumping to judging whether the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set of the wandering reaches the set number of quantitative features of learning rewards, and whether the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge in the knowledge feature binary group where the target cache performance cluster feature set of the wandering is located reaches the set number of cache performance cluster feature sets; when the set wandering complete index is met, the wandering processing is terminated to obtain the set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge.

[0161] The above-mentioned embodiment specifies that the cache performance cluster feature sets in the knowledge feature binary corresponding to each commonality measurement coefficient are walked in descending order of the commonality measurement coefficient to determine the cache performance cluster feature sets corresponding to each of the reinforcement learning decision knowledge to be processed, and the steps of obtaining the setting association mechanism of the cache performance cluster feature sets and the reinforcement learning decision knowledge are further described in detail as follows.

[0162] 1. Basic process of wandering and judgment

[0163] Introduction of walking and judgment logic: When walking the cache performance cluster feature set in descending order of the commonality metric coefficient to determine the corresponding relationship with the reinforcement learning decision knowledge to be processed, additional judgment conditions are introduced. First, walk the cache performance cluster feature set in the learning reward quantization feature pair corresponding to each commonality metric coefficient in descending order. The learning reward quantization feature pair here is a concept associated with the knowledge feature binary mentioned above, which reflects to a certain extent the relationship between the cache performance cluster feature set and the reinforcement learning decision knowledge to be processed in terms of reward quantization.

[0164] During the walking process, it is necessary to determine whether the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set of the walking has reached the set number of learning reward quantization features, and it is also necessary to determine whether the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge in the knowledge feature binary group where the target cache performance cluster feature set of the walking is located has reached the set number of cache performance cluster feature sets.

[0165] Explanation of judgment conditions: Setting the number of learning reward quantization features and setting the number of cache performance cluster feature sets are pre-set thresholds, which are set based on system requirements and optimization goals. For example, the number of learning reward quantization features is set to 3, and the number of cache performance cluster feature sets is set to 2. This means that during the walking process, when the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set reaches 3, and the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge reaches 2, certain association conditions are met.

[0166] For example, there are multiple knowledge feature binary tuples, such as (decision knowledge 1, feature set 1), (decision knowledge 1, feature set 2), (decision knowledge 2, feature set 3), etc., and their commonality coefficients are calculated to be 0.8, 0.7, 0.6, etc., respectively, and are arranged in descending order as (decision knowledge 1, feature set 1), (decision knowledge 1, feature set 2), (decision knowledge 2, feature set 3). When starting to walk, first consider feature set 1 in the knowledge feature binary tuple (decision knowledge 1, feature set 1) as the target cache performance cluster feature set. Then count the number of reinforcement learning decision knowledge to be processed corresponding to feature set 1, and the number of cache performance cluster feature sets corresponding to decision knowledge 1.

[0167] (II) Operations when the judgment conditions are met

[0168] Determine the corresponding relationship and continue walking: If the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set reaches the set number of learning reward quantization features, and the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge reaches the set number of cache performance cluster feature sets, then the target cache performance cluster feature set is determined as the cache performance cluster feature set corresponding to the target pending reinforcement learning decision knowledge. For example, in the above example, if the number of pending reinforcement learning decision knowledge corresponding to feature set 1 reaches 3, and the number of cache performance cluster feature sets corresponding to decision knowledge 1 reaches 2, then it is determined that feature set 1 corresponds to decision knowledge 1.

[0169] Then continue to walk the cache performance cluster feature set in the next knowledge feature binary. The purpose of continuing to walk here is to determine the correspondence between the pending reinforcement learning decision knowledge and the cache performance cluster feature set in other knowledge feature binary tuples. For example, next walk feature set 3 in the knowledge feature binary (decision knowledge 2, feature set 3), and then repeat the previous judgment steps, that is, judge whether the number of pending reinforcement learning decision knowledge corresponding to feature set 3 reaches the set number of learning reward quantization features, and whether the number of cache performance cluster feature sets corresponding to decision knowledge 2 reaches the set number of cache performance cluster feature sets.

[0170] (III) Operations when the judgment conditions are not met

[0171] Directly continue to walk: If the number of pending reinforcement learning decision knowledge corresponding to the target cache performance cluster feature set does not reach the set number of learning reward quantization features, or the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge does not reach the set number of cache performance cluster feature sets, then directly continue to walk the cache performance cluster feature set in the next knowledge feature binary. For example, when considering (decision knowledge 1, feature set 1), if the number of pending reinforcement learning decision knowledge corresponding to feature set 1 is only 2 (not reaching the set 3), or the number of cache performance cluster feature sets corresponding to decision knowledge 1 is only 1 (not reaching the set 2), then directly walk to the next knowledge feature binary, such as (decision knowledge 1, feature set 2), and then perform the same judgment operation on feature set 2.

[0172] 4. Determination of Roaming Termination and Association Mechanism

[0173] Set the complete walking index and termination condition: When the complete walking index is met, the walking process is terminated to obtain the set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge. The complete walking index can be set to complete the above walking and judgment operations for all knowledge feature binary groups. For example, if there are 5 knowledge feature binary groups, when the 5 knowledge feature binary groups are walked and judged according to the above rules, the complete walking index is met.

[0174] Through such a wandering and judgment process, the one-to-one or many-to-one mapping relationship between the cache performance cluster feature set and the reinforcement learning decision knowledge is finally determined, that is, the association mechanism is set. This association mechanism plays an important role in the subsequent determination of the cache resource allocation strategy. For example, in the Internet / supply chain finance scenario, when the cache performance cluster feature set presents a certain state (such as high cache fill rate, high access frequency of specific data types, etc.), according to the set association mechanism, the corresponding reinforcement learning decision knowledge can be quickly determined, so as to select the appropriate cache resource allocation strategy, such as giving priority to caching certain key data (such as transaction order data, supplier credit data, etc.) or adopting a specific cache replacement strategy.

[0175] It can be seen that the beneficial effects of the above embodiments are mainly reflected in the optimization of the association determination process between the cache performance cluster feature set and the reinforcement learning decision knowledge. By introducing additional judgment conditions and wandering logic, the mapping relationship between the two can be established more accurately. In the Internet / supply chain finance scenario, this precise association helps to more effectively utilize reinforcement learning decision knowledge to optimize cache resource allocation. For example, when the cache status of the platform (such as cache fill rate, data access mode, etc.) changes, it can quickly select the appropriate cache resource allocation strategy based on the accurate association mechanism, improve the cache hit rate, reduce the waste of cache resources and system response delays caused by incorrect strategy selection, and thus improve the performance and efficiency of the entire online trading platform.

[0176] In an exemplary embodiment, the obtaining of the second cache performance statistics of several online trading platforms to be processed, before obtaining the several cache performance statistics to be processed, also includes: obtaining the cache optimization feedback quality scores corresponding to several pre-screened online trading platforms respectively; selecting a set proportion of pre-screened online trading platforms from the several pre-screened online trading platforms in descending order of the cache optimization feedback quality scores as the online trading platforms to be processed, to obtain several online trading platforms to be processed.

[0177] The expanded introduction of the above exemplary embodiments can be achieved through the following two aspects.

[0178] 1. Obtaining cache optimization feedback quality scores

[0179] Definition and composition of cache optimization feedback quality score: Before obtaining the second cache performance statistics of several online trading platforms to be processed, it is necessary to first obtain the cache optimization feedback quality scores corresponding to several pre-screened online trading platforms. This cache optimization feedback quality score is a quantitative evaluation of the cache optimization effect of the online trading platform.

[0180] It can be composed of multiple factors, such as the improvement in cache hit rate, the improvement in cache resource utilization, and the contribution to the overall system performance improvement.

[0181] A formula can be defined to calculate the cache optimization feedback quality score:

[0182] (Q=a δH+b δU+c δS);

[0183] Where (Q) represents the cache optimization feedback quality score, (δH) represents the improvement in cache hit rate, (δU) represents the improvement in cache resource utilization, (δS) represents the contribution to the overall system performance improvement (such as the reduction in system response time), (a), (b), and (c) are weight coefficients, which respectively represent the relative importance of the improvement in cache hit rate, the improvement in cache resource utilization, and the contribution to the overall system performance improvement in the cache optimization feedback quality score.

[0184] For example, in the Internet / supply chain finance scenario, for a certain initial screening online trading platform, for example (a=0.4), (b=0.3), (c=0.3). If the cache hit rate of the platform is improved (δH=0.2) (that is, from the previous 30% to 50%), the improvement in cache resource utilization (δU=0.15) (for example, by optimizing the cache strategy, the originally wasted 20% of cache resources is now reduced to 5%), and the contribution to the overall system performance improvement (δS=0.1) (the system response time is shortened from an average of 5 seconds to 4.5 seconds, a reduction of 10%), then according to the formula, it can be calculated:

[0185] (Q=0.4 0.2+0.3 0.15+0.3 0.1=0.08+0.045+0.03=0.155).

[0186] (II) Initial screening of online trading platforms

[0187] Screening by quality score: Screen a set percentage of pre-screened online trading platforms from a number of pre-screened online trading platforms in descending order of cache optimization feedback quality score as online trading platforms to be processed. For example, if there are 100 pre-screened online trading platforms and the percentage is set to 30%, 30 pre-screened online trading platforms will be screened as online trading platforms to be processed.

[0188] First, sort the cache optimization feedback quality scores of each pre-screened online trading platform calculated above. For example, after sorting, the top 30 pre-screened online trading platforms with the highest quality scores are selected as online trading platforms to be processed. The purpose of this screening process is to focus on those online trading platforms that perform well in cache optimization, thereby improving the effectiveness of the subsequent acquisition of the second cache performance statistics and the entire cache resource allocation strategy determination process.

[0189] From the perspective of the Q-reinforcement learning algorithm, these online trading platforms with better performance are more likely to provide valuable data and decision-making knowledge. Because they have achieved certain results in cache optimization, their cache performance statistics and related decision-making knowledge may be closer to the optimal solution. For example, there may already be some effective cache resource allocation strategies in these pending online trading platforms. By analyzing their second cache performance statistics, these effective strategies can be better promoted to the target online trading platform, or cache resource allocation strategies that are more suitable for the target platform can be mined.

[0190] Impact on subsequent steps: This screening method has an important impact on subsequent steps. When obtaining the second cache performance statistics of the online trading platforms to be processed, since these platforms have been screened, their data is more representative and has reference value. For example, when performing feature clustering, based on these higher-quality second cache performance statistics, a more accurate cache performance clustering feature set can be obtained.

[0191] In the subsequent steps of binding the reinforcement learning decision knowledge to be processed with the cache performance cluster feature set and determining the commonality measurement coefficient, the whole process will be more efficient and accurate due to the high quality of the data source (i.e., the online trading platform to be processed). For example, when determining the commonality measurement coefficient, if the data quality is high, the calculated commonality measurement coefficient will more accurately reflect the true relationship between the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set, thereby improving the accuracy of the set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge, and ultimately helping to determine a more optimized cache resource allocation strategy.

[0192] This design can improve the quality and representativeness of data by obtaining cache optimization feedback quality scores and screening out online trading platforms with better performance as online trading platforms to be processed. In the Internet / supply chain finance scenario, this helps to analyze cache performance more accurately, such as more accurately clustering features and more reasonably determining the association between reinforcement learning decision knowledge and cache performance cluster feature sets. This can help formulate more effective cache resource allocation strategies, improve cache hit rates, optimize cache resource utilization, reduce system performance degradation caused by unreasonable cache strategies, and improve the operating efficiency of the entire online trading platform.

[0193] In an expandable embodiment, after determining a first current cache resource allocation strategy from the candidate cache resource allocation strategies based on the relevance adjustment weights of the candidate cache resource allocation strategies, the method further includes: storing the first current cache resource allocation strategy in association with the first cache performance statistics data.

[0194] In detail, the first current cache resource allocation strategy is associated with the first cache performance statistical data and stored, including: obtaining a policy node response logic information sequence according to the first current cache resource allocation strategy, wherein the policy node response logic information sequence includes P groups of policy node response logic information associated before and after, and P is an integer greater than or equal to 1; obtaining a statistical element description information sequence from the first cache performance statistical data according to the policy node response logic information sequence, wherein the statistical element description information sequence includes P groups of statistical element description information associated before and after; based on the policy node response logic information sequence, obtaining a policy node semantic coding sequence through a first semantic coding identification branch included in an associated storage processing algorithm, wherein the policy node semantic coding sequence includes P policy node semantics encoding; based on the statistical element description information sequence, obtaining a statistical element semantic coding sequence through the second semantic coding identification branch included in the associative storage processing algorithm, wherein the statistical element semantic coding sequence includes P statistical element semantic codes; based on the policy node semantic coding sequence and the statistical element semantic coding sequence, obtaining the positive semantic association feature corresponding to the policy node response logic information sequence through the bidirectional long short-term memory branch included in the associative storage processing algorithm; determining the reverse semantic association feature of the policy node response logic information sequence relative to the statistical element description information sequence according to the positive semantic association feature; and associatively storing the first current cache resource allocation strategy with the first cache performance statistical data in combination with the positive semantic association feature and the reverse semantic association feature.

[0195] Furthermore, the method of obtaining the positive semantic association features corresponding to the policy node response logic information sequence based on the policy node semantic coding sequence and the statistical element semantic coding sequence through the bidirectional long short-term memory branch included in the associative storage processing algorithm includes: obtaining P first entity quantitative knowledge based on the policy node semantic coding sequence through the first entity extraction layer included in the associative storage processing algorithm, wherein each first entity quantitative knowledge corresponds to a policy node semantic coding; obtaining P second entity quantitative knowledge based on the statistical element semantic coding sequence through the second entity extraction layer included in the associative storage processing algorithm, wherein each second entity quantitative knowledge corresponds to a statistical element semantic coding; combining the P first entity quantitative knowledge and the P second entity quantitative knowledge to obtain P target entity quantitative knowledge, wherein each target entity quantitative knowledge includes a first entity quantitative knowledge and a second entity quantitative knowledge; obtaining the positive semantic association features corresponding to the policy node response logic information sequence through the bidirectional long short-term memory branch included in the associative storage processing algorithm based on the P target entity quantitative knowledge.

[0196] Furthermore, based on the policy node semantic coding sequence, P first entity quantitative knowledge is obtained through the first entity extraction layer included in the associative storage processing algorithm, including: for each group of policy node semantic coding in the policy node semantic coding sequence, a first pyramid semantic coding is obtained through the pyramid layer included in the first entity extraction layer, wherein the first entity extraction layer belongs to the associative storage processing algorithm; for each group of policy node semantic coding in the policy node semantic coding sequence, a first hole convolutional semantic coding is obtained through the hole convolutional layer included in the first entity extraction layer; for each group of policy node semantic coding in the policy node semantic coding sequence, based on the first pyramid semantic coding and the first hole convolutional semantic coding, a first interaction semantic coding is obtained through the feature interaction layer included in the first entity extraction layer; for each group of policy node semantic coding in the policy node semantic coding sequence, based on the first interaction semantic coding and the policy node semantic coding, a first entity quantitative knowledge is obtained through the first hole convolutional layer included in the first entity extraction layer.

[0197] In actual application, an exemplary technical solution for the above-mentioned associative storage is introduced as follows.

[0198] 1. Overall process of associated storage

[0199] Associative storage of the first current cache resource allocation strategy and the first cache performance statistics: After determining the first current cache resource allocation strategy based on the correlation adjustment weight of the candidate cache resource allocation strategies, the first current cache resource allocation strategy and the first cache performance statistics are associated and stored. The purpose of this associative storage is to establish a logical connection between the cache resource allocation strategy and the corresponding cache performance statistics, so that the information can be better used for analysis, optimization and decision-making in subsequent operations.

[0200] Specifically, this association storage process involves multiple steps. First, a policy node response logic information sequence is obtained according to the first current cache resource allocation policy.

[0201] 2. Acquisition of the strategy node response logic information sequence

[0202] The composition of the strategy node response logic information sequence: The strategy node response logic information sequence includes the (P) groups of strategy node response logic information that are associated with each other ((P) is an integer greater than or equal to (1)). This sequence reflects the response of each logic node during the execution of the first current cache resource allocation strategy. For example, in the Internet / supply chain finance scenario, if the first current cache resource allocation strategy is to prioritize caching transaction order data, then the strategy node response logic information may include the response of the transaction order data cache at different times (such as peak trading period and trough period). For example (P=3), these three groups of strategy node response logic information may correspond to the cache response status during the peak trading period, normal period and trough period, respectively.

[0203] During peak periods, the policy node response logic information may include changes in cache hit rate, impact on system response time, etc. During normal periods, it may include changes in cache resource occupancy ratio, etc. During low periods, it may include the effect of data prefetching strategy, etc. This information constitutes a complete policy node response logic information sequence, describing the execution of cache resource allocation strategy from different time or state perspectives.

[0204] 3. Acquisition of the information sequence describing statistical elements

[0205] Acquisition based on the policy node response logic information sequence: Acquiring the statistical element description information sequence from the first cache performance statistical data according to the policy node response logic information sequence. Since the policy node response logic information sequence reflects the execution logic of the cache resource allocation policy, the statistical element description information related to the first cache performance statistical data can be screened out through this logic.

[0206] The statistical element description information sequence includes (P) groups of statistical element description information that are related to each other. For example, when the policy node response logic information focuses on the change of cache hit rate during the peak period, the corresponding statistical element description information may be the actual cache hit rate value during this period, the access frequency of related data types (such as transaction order data), etc. Similarly, when (P=3), these three groups of statistical element description information correspond to the three groups of information in the policy node response logic information sequence, providing information related to policy execution from the perspective of cache performance statistics.

[0207] 4. Acquisition of semantic coding sequences of strategy nodes and statistical elements

[0208] Acquisition of the semantic coding sequence of the policy node: Based on the policy node response logic information sequence, the policy node semantic coding sequence is acquired through the first semantic coding identification branch included in the associative storage processing algorithm. This semantic coding sequence includes (P) policy node semantic codes. The purpose of semantic coding is to convert the policy node response logic information into a computer-processable coding form for subsequent semantic association analysis.

[0209] For example, a hash coding-based approach can be used. For example, the policy node response logic information during the peak period includes information such as the cache hit rate increased from (60%) to (70%), and the system response time was shortened from an average of (5) seconds to (4) seconds. This information can be converted into a hash value as a policy node semantic code according to certain rules. For example, for the improvement of the hit rate, a function (h_1=hash((60,70))) can be set, and for the shortening of the system response time, (h_2=hash((5,4))) can be set, and then (h_1) and (h_2) can be combined as the policy node semantic code corresponding to the peak period. The same approach can be used for normal and low periods, and finally a policy node semantic code sequence containing (P=3) policy node semantic codes is obtained.

[0210] Acquisition of the semantic coding sequence of statistical elements: Based on the statistical element description information sequence, the statistical element semantic coding sequence is obtained through the second semantic coding identification branch included in the associative storage processing algorithm. This sequence also includes (P) statistical element semantic codes. For example, a similar hash coding method is also used for the cache hit rate value, data access frequency, etc. in the statistical element description information. For example, the cache hit rate value corresponding to the peak period is (70%), and the data access frequency is (100) times per hour, which can be converted into a hash value (h_3=hash((70, 100))), which is part of the statistical element semantic coding corresponding to the peak period. In this way, each group of statistical element description information is encoded, and finally a statistical element semantic coding sequence is obtained.

[0211] 5. Acquisition of positive semantic association features

[0212] Acquisition of the first entity quantitative knowledge and the second entity quantitative knowledge: Based on the semantic coding sequence of the policy node and the semantic coding sequence of the statistical element, the positive semantic association features corresponding to the response logic information sequence of the policy node are acquired through the bidirectional long short-term memory branch included in the associative storage processing algorithm. First, based on the semantic coding sequence of the policy node, (P) first entity quantitative knowledge are acquired through the first entity extraction layer included in the associative storage processing algorithm, where each first entity quantitative knowledge corresponds to a policy node semantic coding.

[0213] Based on each group of policy node semantic codes in the policy node semantic code sequence, the first pyramid semantic code is obtained through the pyramid layer included in the first entity refining layer. For example, for the policy node semantic code corresponding to the peak period mentioned above, the pyramid layer can perform multi-scale feature extraction on it and convert it into the first pyramid semantic code. For example, an exemplary pyramid layer algorithm is used to decompose and combine the original policy node semantic code according to different resolutions to obtain a multi-scale coding form.

[0214] At the same time, for each group of policy node semantic codes in the policy node semantic code sequence, the first dilated convolutional semantic code is obtained through the dilated convolutional layer included in the first entity extraction layer. The dilated convolutional layer can expand the receptive field of the convolution kernel without increasing the amount of calculation, thereby obtaining more comprehensive semantic information. For example, for the peak period policy node semantic code, the first dilated convolutional semantic code is obtained through the dilated convolution operation.

[0215] Then, for each group of policy node semantic codes, based on the first pyramid semantic code and the first dilated convolution semantic code, the first interactive semantic code is obtained through the feature interaction layer included in the first entity extraction layer. The function of the feature interaction layer is to interactively fuse the semantic codes obtained by the pyramid layer and the dilated convolution layer to obtain a more comprehensive semantic code. For example, the first interactive semantic code is obtained by weighted summing or other fusion operations of the first pyramid semantic code and the first dilated convolution semantic code.

[0216] Finally, for each group of policy node semantic codes, based on the first interaction semantic code and the policy node semantic code, the first entity quantitative knowledge is obtained through the first hole convolution layer included in the first entity extraction layer. This first entity quantitative knowledge is a quantitative representation of the policy node semantic code obtained after multi-layer processing, which contains richer semantic information.

[0217] Similarly, based on the statistical element semantic coding sequence, (P) second entity quantitative knowledge is obtained through the second entity extraction layer included in the associative storage processing algorithm, where each second entity quantitative knowledge corresponds to a statistical element semantic coding. This process is similar to obtaining the first entity quantitative knowledge, and the statistical element semantic coding is processed through operations such as the pyramid layer, the hollow convolution layer, and the feature interaction layer to obtain the second entity quantitative knowledge.

[0218] Combination of target entity quantitative knowledge and acquisition of positive semantic association features: Combine (P) first entity quantitative knowledge and (P) second entity quantitative knowledge to obtain (P) target entity quantitative knowledge, where each target entity quantitative knowledge includes a first entity quantitative knowledge and a second entity quantitative knowledge. For example, the first entity quantitative knowledge and the second entity quantitative knowledge corresponding to the peak period are combined to form a target entity quantitative knowledge.

[0219] Based on (P) target entity quantitative knowledge, the positive semantic association features corresponding to the policy node response logic information sequence are obtained through the bidirectional long short-term memory branch included in the associative storage processing algorithm. Bidirectional long short-term memory network (Bi-LSTM) is a deep learning model that can process sequence data. It can process the target entity quantitative knowledge sequence and capture the long-term dependencies in the sequence to obtain positive semantic association features. For example, Bi-LSTM can calculate a feature vector representing positive semantic association based on the order and semantic relationship between each element in the target entity quantitative knowledge sequence. This feature vector reflects the degree of semantic association in the positive direction between the policy node response logic information sequence and the statistical element description information sequence.

[0220] 6. Determination of reverse semantic association features

[0221] Determine the reverse semantic association feature based on the positive semantic association feature: Determine the reverse semantic association feature of the strategy node response logic information sequence relative to the statistical element description information sequence based on the positive semantic association feature. In the semantic association analysis, the positive and reverse semantic association features are interrelated. A symmetry-based or complementary method can be used to determine the reverse semantic association feature.

[0222] For example, if the positive semantic association feature is represented by a feature vector (vec{F}), then the reverse semantic association feature can be obtained by performing some transformation on (vec{F}). An exemplary method is to invert or complement the elements in (vec{F}) according to certain rules. For example, (vec{F}=(f_1, f_2, f_3)), then the reverse semantic association feature vector (vec{F}'=(1-f_1, 1-f_2, 1-f_3)) (This is just an example, the actual calculation method may be more complicated, depending on the specific algorithm design).

[0223] (VII) Final Implementation of Associative Storage

[0224] Associative storage combining positive and negative semantic association features: The first current cache resource allocation strategy and the first cache performance statistics data are associated and stored by combining positive semantic association features and negative semantic association features. This associative storage method not only considers the positive association between the strategy and the performance data, but also considers the negative association, thereby more comprehensively describing the relationship between them.

[0225] For example, the forward and reverse semantic association features can be stored in a database as metadata together with the first current cache resource allocation strategy and the first cache performance statistics data. When the relationship between the cache resource allocation strategy and the performance data needs to be queried or analyzed later, more accurate retrieval and analysis can be performed based on these semantic association features.

[0226] In this way, by associating and storing the first current cache resource allocation strategy with the first cache performance statistics data, and using complex semantic encoding, entity extraction, and bidirectional long and short-term memory technologies to construct positive and negative semantic association features, the relationship between the strategy and performance data can be fully and deeply described. In the Internet / supply chain finance scenario, this associative storage helps to better understand the impact of cache resource allocation strategies on cache performance under different circumstances (such as peak trading periods, trough periods, etc.). It can provide richer reference information for subsequent cache strategy optimization, improve cache hit rate, optimize cache resource utilization, and improve the operating efficiency of the entire online trading platform. It also provides more effective data support for troubleshooting, performance analysis, etc.

[0227] Further, Figure 2 FIG. 2 is a schematic diagram of a platform cache performance optimization system 200 provided in an embodiment of the present invention. Figure 2 The platform cache performance optimization system 200 shown includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method in the embodiment of the present invention.

[0228] Alternatively, if Figure 2As shown, the platform cache performance optimization system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention.

[0229] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0230] Alternatively, if Figure 2 As shown, the platform cache performance optimization system 200 may further include a transceiver 220. The processor 210 may control the transceiver 220 to interact with other devices. Specifically, the transceiver 220 may send information or data to other devices, or receive information or data sent by other devices.

[0231] Optionally, the platform cache performance optimization system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present invention, which will not be described in detail here for the sake of brevity.

[0232] It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capability.

[0233] It is understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0234] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0235] It should be noted that, in this article, the term "comprises", "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0236] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned implementation methods can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the embodiment of the present invention.

[0237] The embodiments of the embodiments of the present invention are described above in conjunction with the accompanying drawings, but the embodiments of the present invention are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present invention and the scope of protection of the embodiments of the present invention, all of which are within the protection of the embodiments of the present invention.

Claims

1. A method for improving the cache performance of an online trading platform, characterized in that: The method is applied to a platform cache performance optimization system, and the method comprises: Acquire first cache performance statistics of the target online trading platform; wherein the first cache performance statistics include data access status information, cache resource status data and cache policy response information; Determine a candidate cache resource allocation strategy that pairs the first reinforcement learning decision knowledge with the first cache performance statistical data, and an initial correlation weight of the candidate cache resource allocation strategy; wherein the first reinforcement learning decision knowledge is obtained by learning the associated cache performance statistical data of an associated online trading platform through a Q-reinforcement learning algorithm, and the associated online trading platform and the target online trading platform belong to the same online trading scenario; Determining adjustment characteristics of the candidate cache resource allocation strategy based on the timeliness requirement level of the candidate cache resource allocation strategy; Adjusting the initial relevance weight of the candidate cache resource allocation strategy based on the adjustment characteristics of the candidate cache resource allocation strategy to obtain the relevance adjustment weight of the candidate cache resource allocation strategy; determining a first current cache resource allocation strategy from the candidate cache resource allocation strategies based on the relevance adjustment weights of the candidate cache resource allocation strategies; The determining, based on the timeliness requirement level of the candidate cache resource allocation strategy, the adjustment characteristics of the candidate cache resource allocation strategy includes: Determining the timeliness requirement level as the current cache performance optimization decision weight of the candidate cache resource allocation strategy; Obtaining past policy execution trust weights and past cache performance optimization decision weights of the candidate cache resource allocation policy; Determining a current policy execution trust weight of the candidate cache resource allocation policy based on the timeliness requirement level and the past policy execution trust weight; Determining a current platform mode matching weight of the candidate cache resource allocation strategy based on the timeliness requirement level and the past cache performance optimization decision weight; Based on the current cache performance optimization decision weight, the current policy execution trust weight and the current platform pattern matching weight of the candidate cache resource allocation strategy, the adjustment characteristics of the candidate cache resource allocation strategy are determined.

2. The method for improving the cache performance of an online trading platform according to claim 1, characterized in that: Before determining the timeliness requirement level as the current cache performance optimization decision weight of the candidate cache resource allocation strategy, the method further includes: Obtain the association mechanism between the online transaction session period and the cache time tolerance; Based on the set association mechanism, determine the target cache time tolerance corresponding to the current online transaction session period; Obtaining a cache timeliness warning score for the candidate cache resource allocation strategy; Based on the target cache timeliness tolerance and the cache timeliness warning score, a timeliness requirement level of the candidate cache resource allocation strategy is determined.

3. The method for improving the cache performance of an online trading platform according to claim 2, characterized in that: The obtaining of the past policy execution trust weights and past cache performance optimization decision weights of the candidate cache resource allocation policy includes: Based on the set association mechanism, determining the past target cache time tolerance corresponding to the past online transaction session period; Obtaining past cache timeliness warning scores of the candidate cache resource allocation strategy; Determining a past timeliness requirement level of the candidate cache resource allocation strategy based on the past target cache timeliness tolerance and the past cache timeliness warning score; Based on the past timeliness requirement levels, past policy execution trust weights and past cache performance optimization decision weights of the candidate cache resource allocation policies are determined.

4. The method for improving the cache performance of an online trading platform according to any one of claims 1 to 3, characterized in that: After determining the first current cache resource allocation strategy from the candidate cache resource allocation strategies based on the correlation adjustment weights of the candidate cache resource allocation strategies, the method further includes: Acquire second cache performance statistics of the target online trading platform; Determine a target cache performance cluster feature set corresponding to the second cache performance statistical data based on a set association mechanism between cache performance statistical data and cache performance cluster feature sets, wherein the cache performance cluster feature set is obtained by performing feature clustering on the cache performance statistical data corresponding thereto; Determine the target learning reward quantitative feature corresponding to the target cache performance cluster feature set based on the set association mechanism between the cache performance cluster feature set and the learning reward quantitative feature; Determine the resource allocation optimization strategy corresponding to the target learning reward quantization feature as the second current cache resource allocation strategy; Determine a third current cache resource allocation strategy from the first current cache resource allocation strategy and the second current cache resource allocation strategy; The third current cache resource allocation strategy is recommended to the target online trading platform.

5. The method for improving the cache performance of an online trading platform as claimed in claim 4, characterized in that: Before obtaining the second cache performance statistics of the target online trading platform, the method further includes: Obtaining second cache performance statistics of a plurality of online trading platforms to be processed, and obtaining a plurality of cache performance statistics to be processed; Performing feature clustering on a plurality of cache performance statistical data to be processed to obtain a plurality of cache performance cluster feature sets; The cache performance statistics data to be processed of each cache performance cluster feature set obtained by the feature cluster are determined as the cache performance statistics data to be processed corresponding to each cache performance cluster feature set, and a setting association mechanism between the cache performance statistics data and the cache performance cluster feature set is obtained.

6. The method for improving the cache performance of an online trading platform according to claim 5, characterized in that: After performing feature grouping on the plurality of cache performance statistical data to be processed to obtain a plurality of cache performance group feature sets, the method further includes: Obtaining a number of reinforcement learning decision-making knowledge to be processed; Binding a plurality of to-be-processed reinforcement learning decision knowledge with a plurality of cache performance cluster feature sets one by one to obtain a plurality of knowledge feature binary tuples, each of the knowledge feature binary tuples comprising a to-be-processed reinforcement learning decision knowledge and a cache performance cluster feature set; Determine the commonality measurement coefficient of the reinforcement learning decision knowledge to be processed and the cache performance cluster feature set in each of the knowledge feature binary groups, and obtain the commonality measurement coefficient corresponding to each of the knowledge feature binary groups; The cache performance cluster feature sets in the knowledge feature binary corresponding to each commonality measurement coefficient are walked in descending order to determine the cache performance cluster feature sets corresponding to each of the reinforcement learning decision knowledge to be processed, and the set association mechanism between the cache performance cluster feature sets and the reinforcement learning decision knowledge is obtained.

7. The method for improving the cache performance of an online trading platform according to claim 6, characterized in that: The cache performance cluster feature set in the knowledge feature binary corresponding to each commonality measurement coefficient is walked in descending order to determine the cache performance cluster feature set corresponding to each of the reinforcement learning decision knowledge to be processed, and the set association mechanism of the cache performance cluster feature set and the reinforcement learning decision knowledge is obtained, including: Walk the cache performance cluster feature sets in the learning reward quantization feature pairs corresponding to each commonality metric coefficient in descending order, and determine whether the number of to-be-processed reinforcement learning decision knowledge corresponding to the wandering target cache performance cluster feature set reaches the set number of learning reward quantization features, and whether the number of cache performance cluster feature sets corresponding to the target to-be-processed reinforcement learning decision knowledge in the knowledge feature binary group where the wandering target cache performance cluster feature set is located reaches the set number of cache performance cluster feature sets; If so, the target cache performance cluster feature set is determined as the cache performance cluster feature set corresponding to the target to-be-processed reinforcement learning decision knowledge, and the cache performance cluster feature set in the next knowledge feature binary group is continued to be wandered, and the process is jumped to the step of judging whether the number of to-be-processed reinforcement learning decision knowledge corresponding to the wandering target cache performance cluster feature set reaches the set number of learning reward quantization features, and whether the number of cache performance cluster feature sets corresponding to the target to-be-processed reinforcement learning decision knowledge in the knowledge feature binary group where the wandering target cache performance cluster feature set is located reaches the set number of cache performance cluster feature sets; If not, continue to wander the cache performance cluster feature set in the next knowledge feature binary, and jump to the step of judging whether the number of pending reinforcement learning decision knowledge corresponding to the wandering target cache performance cluster feature set reaches the set number of learning reward quantization features, and whether the number of cache performance cluster feature sets corresponding to the target pending reinforcement learning decision knowledge in the knowledge feature binary where the wandering target cache performance cluster feature set is located reaches the set number of cache performance cluster feature sets; When the set walking completeness index is met, the walking process is terminated, and the set association mechanism between the cache performance cluster feature set and the reinforcement learning decision knowledge is obtained.

8. The method for improving the cache performance of an online trading platform according to claim 5, characterized in that: The step of obtaining the second cache performance statistics of the plurality of online trading platforms to be processed, before obtaining the plurality of cache performance statistics to be processed, further includes: Obtain cache optimization feedback quality scores corresponding to several initially screened online trading platforms; A set proportion of initially screened online trading platforms is selected from the plurality of initially screened online trading platforms in descending order of the cache optimization feedback quality score as online trading platforms to be processed, thereby obtaining a plurality of online trading platforms to be processed.

9. A platform cache performance optimization system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 8.

Images