A hotel revenue management decision method and system based on machine learning
By constructing decision state vectors and evaluating candidate actions using Bayesian neural networks, the problems of data processing lag and pricing strategy lag in traditional hotel revenue management are solved, enabling real-time response to market changes and dynamic optimization of revenue.
Patent Information
- Application Number
- CN202610565051.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-27
- Publication Date
- 2026-08-04
AI Technical Summary
Traditional hotel revenue management methods struggle to handle massive amounts of heterogeneous data from multiple sources, are unable to respond to market changes in real time, result in lagging pricing strategies, and fail to achieve the optimal revenue trade-off in non-steady-state environments.
By constructing a decision state vector to quantify environmental uncertainty, using a Bayesian neural network to evaluate the expected returns and uncertainties of candidate actions, and combining this with a risk-constrained objective function to optimize inventory control, dynamic and adaptive revenue management decisions are achieved.
It enables accurate perception of market instability, adaptive adjustment of pricing and inventory strategies, avoidance of revenue fluctuations, and improvement of the flexibility and accuracy of hotel revenue management.
Smart Images
Figure CN122509948A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hotel revenue management technology, specifically to a hotel revenue management decision-making method and system based on machine learning. Background Technology
[0002] With the acceleration of globalization and the booming development of business and leisure tourism markets, the hotel industry, as a crucial pillar of modern service industries, faces unprecedented opportunities and challenges. Revenue management, as a core strategy of hotel operations, essentially maximizes room revenue and overall profit by selling the right products to the right customers at the right prices through scientific pricing and inventory control, based on market demand forecasting. Traditional hotel revenue management relies primarily on the personal experience of revenue managers, simple linear regression analysis, and static reviews of historical booking data. This model played a significant role during periods of relatively stable market conditions and low information asymmetry. However, in today's digital economy era, with rapidly changing market supply and demand and increasingly complex consumer behavior, traditional methods have shown significant limitations in handling massive data processing, real-time market response, and multi-dimensional variable correlation analysis.
[0003] The core pain points of traditional revenue management methods lie in their lag and subjectivity. First, manual analysis struggles to effectively handle massive amounts of heterogeneous data from multiple sources, including hotel management systems, online travel agencies, search engines, social media, and macroeconomic indicators. This data is often characterized by high dimensionality, non-linearity, and spatiotemporal correlation, making manual cleaning, integration, and insight nearly impossible. Second, traditional statistical models typically assume a stable market environment, failing to capture the sudden impact of unforeseen events (such as large-scale exhibitions, holidays, public health emergencies, or extreme weather) on demand, significantly reducing the accuracy of demand forecasts. Furthermore, traditional pricing strategies are often based on fixed room types and packages, lacking flexibility and unable to dynamically adjust to real-time booking progress, competitor price changes, and customer price sensitivity within milliseconds. This "one-size-fits-all" or "empirical" pricing approach often results in a failure to achieve price premiums during peak demand periods or the loss of potential customers during off-peak periods due to excessively high prices, ultimately leading to revenue loss.
[0004] Chinese invention patent CN120338035A discloses a multi-agent collaborative method based on a large language model, comprising: constructing a multi-agent system; establishing a multi-agent coordinated resource allocation model; transforming the task allocation problem in the resource allocation model into a multi-objective optimization problem, with optimization objectives including minimizing task completion time and optimizing resource utilization; and using an improved particle swarm optimization (PSO) algorithm to solve the multi-objective optimization problem to obtain the optimal collaborative scheme among the master agent, task creation agent, intent orchestration agent, and at least one specialized agent. The improved PSO algorithm evaluates the uncertainty of each candidate solution through multiple random samplings, and when the uncertainty exceeds a threshold, cross-validation is performed through a multi-agent voting mechanism. Addressing the accumulation of uncertainty in multi-agent systems under complex cognitive tasks, this application improves the efficiency and stability of collaborative decision-making and task scheduling in multi-agent systems.
[0005] However, the above and similar technical solutions still have the following shortcomings: In scenarios such as dynamic pricing, although strategies such as "small price adjustments" can be adopted to respond and adjust based on real-time market feedback, the market environment is time-varying and consumer preferences and competitive landscape are constantly evolving. Therefore, in non-steady-state environments, it is impossible to provide corresponding intelligent support for key decisions such as dynamic pricing and inventory allocation, making it difficult to achieve the optimal trade-off between "exploring the potential benefits of unknown price ranges" and "using the known optimal price to ensure current benefits". Summary of the Invention
[0006] The purpose of this invention is to provide a hotel revenue management decision-making method and system based on machine learning to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a hotel revenue management decision-making method based on machine learning, comprising:
[0008] S1: Quantification of uncertainty in environmental perception: Construct a decision state vector corresponding to the current market state through multi-source data from inside and outside the hotel, and determine the corresponding environmental uncertainty metric value based on the volatility of the decision state vector within a preset time window.
[0009] S2: Dynamic generation of exploration-utilization strategy: Based on the decision state vector and the set value assessment model, obtain the expected total revenue valuation and revenue uncertainty valuation corresponding to each candidate action in the revenue management action list, and determine the corresponding candidate action to be executed through the environmental uncertainty metric.
[0010] S3: Risk-constrained strategy execution: Based on the candidate actions to be executed and the decision state vector, determine the corresponding expected demand and probability distribution, construct the corresponding objective function, and obtain the corresponding inventory control instructions through the objective function.
[0011] Furthermore, the corresponding environmental uncertainty measures are determined, including:
[0012] S1.1: Vector Construction: Collect and obtain corresponding internal and external multi-source data through internal systems, third-party data services and event calendars, and obtain corresponding pre-processed multi-source data through data cleaning and standardization. At the same time, determine the corresponding static features, dynamic market features, time series features and event features based on the pre-processed multi-source data, and concatenate the static features, dynamic market features, time series features and event features to obtain the corresponding decision state vector.
[0013] S1.2: Value Determination: By collecting and acquiring all decision state vectors within a preset review time window, all corresponding dynamic market characteristics within the preset review time window are obtained, the corresponding dynamic market characteristic standard deviation is determined, and the corresponding normalized dynamic market characteristic standard deviation is obtained through a preset scaling function. At the same time, the corresponding environmental uncertainty metric is determined based on the normalized dynamic market characteristic standard deviation.
[0014] Furthermore, the dynamic market characteristics include price dimension data, demand dimension data, and booking rate dimension data. Based on the preset scaling function, the standard deviations of the price dimension data, demand dimension data, and booking rate dimension data are normalized to obtain the corresponding volatility components of the price dimension, demand dimension, and booking rate dimension. Simultaneously, these volatility components are combined to determine the corresponding environmental uncertainty measure, specifically:
[0015]
[0016] in: This is a measure of environmental uncertainty. For the volatility component in the price dimension, As a volatility component of the demand dimension, For the volatility component of the booking rate dimension, , , For component weights.
[0017] Furthermore, the corresponding candidate actions to be executed are identified, including:
[0018] S2.1: Model building: By constructing training data pairs based on the daily market status, daily decision-making actions, and daily business revenue generated from the hotel's historical data, a Bayesian neural network is trained using the training data pairs to obtain the corresponding value assessment model.
[0019] S2.2: Action Evaluation: Based on the hotel's corresponding revenue management operations, determine the corresponding pricing actions and inventory control actions, set up the corresponding revenue management action list, and pair the decision state vector with each candidate action in the revenue management action list one by one to construct the corresponding vector action data pair. The vector action data pair is used as the input of the value evaluation model, and the output is to obtain the corresponding expected total revenue valuation and revenue uncertainty valuation.
[0020] S2.3: Decision Determination: Based on the strategy parameters set according to the value assessment model, the expected total revenue valuation and revenue uncertainty valuation are processed, and based on the processed expected total revenue valuation and revenue uncertainty valuation, the corresponding candidate actions to be executed are determined from the revenue management action list.
[0021] Furthermore, the value assessment model includes a strategy configuration module, which sets corresponding strategy parameters based on the behavioral algorithm employed by the value assessment model, specifically:
[0022] When the behavior algorithm is an uncertainty-driven ε-greedy policy, the corresponding policy parameters include the base exploration probability and the uncertainty amplification coefficient; when the behavior algorithm is a UCB policy that incorporates environmental uncertainty, the corresponding policy parameters include the exploration weight coefficient.
[0023] Furthermore, when the strategy parameters include a base exploration probability and an uncertainty amplification coefficient, the base exploration probability and the environmental uncertainty metric are combined to obtain the corresponding exploration probability. A corresponding uniform random number is then set using the pseudo-random number generator built into the value assessment model. Simultaneously, based on the comparison between the exploration probability and the uniform random number, the corresponding execution logic is determined, specifically as follows:
[0024] When the exploration probability is greater than a uniform random number, the corresponding execution logic is the exploitation mode; otherwise, the corresponding execution logic is the exploration mode.
[0025] Furthermore, when the execution logic is in the utilization mode, the expected total revenue valuations corresponding to each candidate action are compared to determine the maximum expected total revenue valuation, and the candidate action corresponding to the maximum expected total revenue valuation is set as the candidate action to be executed.
[0026] When the execution logic is in exploration mode, the median of the revenue uncertainty estimate is determined based on the revenue uncertainty estimate corresponding to each candidate action. A high-uncertainty action candidate pool is constructed using revenue uncertainty estimates higher than the median revenue uncertainty estimate. The sum of the uncertainty estimates of the candidate pool is obtained. Simultaneously, the selection probability corresponding to each revenue uncertainty estimate in the high-uncertainty action candidate pool is set based on the ratio between the revenue uncertainty estimate and the sum of the uncertainty estimates of the candidate pool. Weighted random sampling is performed based on the selection probability, and the candidate action corresponding to the revenue uncertainty estimate of the weighted random sampling result is set as the corresponding candidate action to be executed.
[0027] Furthermore, when the strategy parameters include an exploration weight coefficient, the exploration weight coefficient, the environmental uncertainty metric, and the return uncertainty estimate are combined to obtain the corresponding uncertainty potential value. The uncertainty potential value is then combined with the expected total return estimate to obtain the corresponding comprehensive UCB score. Simultaneously, the comprehensive UCB scores corresponding to each candidate action are compared to determine the highest comprehensive UCB score. The candidate action corresponding to the highest comprehensive UCB score is then set as the corresponding candidate action to be executed.
[0028] Furthermore, obtain the corresponding inventory control instructions, including:
[0029] S3.1: Model prediction: By constructing training sample data pairs based on the historical states, actions, and results corresponding to the hotel's historical data, a probabilistic prediction model is trained using the training sample data pairs to obtain a demand prediction model. At the same time, the decision state vector and the candidate actions to be executed are both used as inputs to the demand prediction model, and the corresponding expected demand and probability distribution are output.
[0030] S3.2: Function Construction: Based on the actual number of rooms sold and the room price for each room type in the hotel, determine the corresponding current immediate revenue value. Combine the current immediate revenue value, expected demand, and probability distribution to set the corresponding objective function. Simultaneously, optimize the objective function using a back induction algorithm to obtain the optimal value function. Combine the remaining room inventory and remaining stay days corresponding to the optimal value function with the expected demand and probability distribution to determine the corresponding final reserved inventory and current available quota. At the same time, construct the corresponding inventory control instruction based on the room price, final reserved inventory, and current available quota corresponding to the candidate action to be executed.
[0031] A machine learning-based hotel revenue management decision system uses any one of the machine learning-based hotel revenue management decision methods described above.
[0032] Compared with the prior art, the beneficial effects of the present invention are:
[0033] Firstly, this invention integrates multi-source data from both inside and outside the hotel to construct a corresponding decision state vector. Based on market volatility within a preset time window, it calculates the corresponding environmental uncertainty metric, thereby achieving dynamic and accurate perception of the degree of market instability. This effectively overcomes the shortcomings of traditional revenue management methods that rely on lagging or static data and are difficult to keenly capture drastic market fluctuations, providing a high-fidelity environmental benchmark for dynamic decision-making.
[0034] Secondly, this invention uses a Bayesian neural network to conduct a dual evaluation of expected returns and uncertainties for candidate pricing and inventory actions, and performs adaptive adjustment of the exploration-utilization strategy. This enables the automatic weighing of risks and opportunities in a complex and ever-changing market, avoiding the rigidity of fixed thresholds or heuristic strategies, and improving the adaptive tuning capability of the strategy.
[0035] Thirdly, this invention outputs the expected demand and its probability distribution by training a demand forecasting model, while constructing an objective function that includes risk aversion factors. It then uses a back-induction algorithm to perform global optimization, taking into account both immediate revenue and long-term value. This allows the hotel to accurately calculate the final inventory level, current available quota, and room pricing based on its own risk preferences, generating inventory control instructions that take into account risk constraints. This effectively avoids significant revenue fluctuations or room vacancies caused by aggressive pricing. Attached Figure Description
[0036] Figure 1 This is a flowchart illustrating the hotel revenue management decision-making method of the present invention;
[0037] Figure 2 This is a flowchart illustrating the uncertainty quantification method for environmental perception in this invention.
[0038] Figure 3 This is a schematic diagram of the processing flow of the value assessment model in this invention;
[0039] Figure 4 This is a schematic diagram of the process for determining candidate actions to be executed when the value assessment model in this invention adopts an uncertainty-driven ε-greedy strategy;
[0040] Figure 5 This is a schematic diagram illustrating the process of determining candidate actions to be executed when the value assessment model in this invention adopts the UCB strategy that incorporates environmental uncertainty. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] refer to Figure 1 This embodiment provides a hotel revenue management decision-making method based on machine learning, which specifically includes the following steps:
[0043] Step S1: Quantification of Uncertainty in Environmental Perception. This involves collecting multi-source data from both inside and outside the hotel to construct a decision state vector corresponding to the current market state. Based on the volatility of the constructed decision state vector within a preset time window, the degree of non-steadiness in the current market is quantified to determine the corresponding environmental uncertainty metric.
[0044] Step S2: Exploration - Dynamic Generation of Utilization Strategies. Based on the decision state vector constructed in Step S1 and the established value assessment model, obtain the expected total revenue estimate and revenue uncertainty estimate for each candidate action in the revenue management action list. Simultaneously, using the environmental uncertainty metric determined in Step S1, dynamically adjust the exploration strategy strength of the value assessment model to identify corresponding candidate actions to be executed from the revenue management action list.
[0045] Step S3: Execution of risk-constrained strategies. This involves using the candidate actions identified in Step S2 and the decision state vector obtained in Step S1 as input to the established demand forecasting model, outputting the corresponding expected demand and probability distribution. Simultaneously, based on the output expected demand and probability distribution, a corresponding objective function is set to obtain the corresponding inventory control instructions.
[0046] This embodiment also provides a hotel revenue management decision system based on machine learning, which uses the aforementioned machine learning-based hotel revenue management decision method.
[0047] In this embodiment, a decision state vector corresponding to the current market state is constructed by real-time acquisition of multi-source data from both inside and outside the hotel. This decision state vector includes, but is not limited to, static hotel attributes (e.g., star rating, city), dynamic market signals (e.g., competitor hotel prices, channel search volume, market supply and demand index), time-series information (e.g., advance notice period, current date), and special events (e.g., holidays, large-scale events). Simultaneously, based on the fluctuation range of the decision state vector, a corresponding environmental uncertainty metric is quantified to determine the current market fluctuation state. (Reference) Figure 2 This embodiment provides a method for quantifying the uncertainty of environmental perception, which specifically includes the following steps:
[0048] Step S1.1: Vector Construction. This involves collecting relevant internal and external multi-source data through internal systems, third-party data services, and event calendars. Specifically, through the hotel's PMS (Property Management System), real-time room status, historical booking records, and the number of rooms sold through various channels are obtained. Through the CRM (Customer Relationship Management) system, relevant customer profile data is obtained, including but not limited to historical spending levels and stay preferences. Simultaneously, through third-party data services, real-time publicly available prices and room status (i.e., booking status) of competing hotel groups are obtained, and through search engines or large online travel agency platforms, relevant market search popularity indices and destination traffic trend data are obtained. Finally, through holiday calendars and official announcement platforms for major local events (such as conventions, sporting events, and concerts), the event calendar corresponding to each time period is determined.
[0049] Furthermore, the collected internal and external multi-source data undergoes data cleaning and standardization processing, including but not limited to handling missing values, correcting outliers, and data alignment (i.e., consistency in time (e.g., standardizing to the "hour" or "day" level) and statistics (e.g., standardizing all prices to "net price including tax"), to obtain the corresponding preprocessed multi-source data. It is worth noting that the data cleaning and standardization processing in this embodiment are conventional techniques, and therefore are not specifically described in this embodiment.
[0050] Furthermore, by acquiring the preprocessed multi-source data, the corresponding static features, dynamic market features, temporal features, and event features are determined. Specifically, the static features in this embodiment are the hotel's own attributes, such as the city category (e.g., first-tier city, tourist city), star rating, total number of rooms, and main customer groups. The dynamic market features in this embodiment include competitive landscape and supply and demand relationship. The competitive landscape includes price dimension data (i.e., the price difference between the hotel's price and the average price of the competing group) and demand dimension data (i.e., the price difference between the hotel's price and the lowest price of the competing group). The supply and demand relationship includes booking rate dimension data (i.e., the ratio between the number of available rooms for the expected target date and the total number of rooms). The temporal features in this embodiment are the periodic codes corresponding to the event calendar (e.g., sine / cosine codes), and the corresponding advance booking days are determined based on the number of days between the check-in date and the booking date. In this embodiment, the event features are assigned corresponding influence intensity coefficients based on the known events (such as holidays or large-scale events). These coefficients can be specifically set according to actual needs, so they are not specifically described in this embodiment. For example, if the known event is a regular weekend, the influence intensity coefficient is set to 1; if the known event is the National Day Golden Week, the influence intensity coefficient is set to 5; and if the known event is a mega-concert, the influence intensity coefficient is set to 3. In other words, based on the determined static features, dynamic market features, time-series features, and event features, features are concatenated according to a preset concatenation order to obtain the corresponding decision state vector.
[0051] Step S1.2: Value Determination. This involves collecting all decision state vectors within a preset review time window (which can be specifically set according to actual needs, so this embodiment does not elaborate on it specifically, e.g., the past 7 days and the past 30 days), and obtaining all dynamic market characteristics corresponding to the preset review time window using these vectors. Simultaneously, based on all dynamic market characteristics within the preset review time window, the corresponding standard deviation of the dynamic market characteristics is obtained. This standard deviation is then normalized using a preset scaling function (which can be specifically set according to actual needs, so this embodiment does not elaborate on it specifically, e.g., dividing by the maximum value of the corresponding dynamic market characteristic within the preset review time window, or using logarithmic scaling), to obtain the corresponding normalized standard deviation of the dynamic market characteristics.
[0052] Specifically, based on all decision state vectors within a preset review time window, all corresponding price, demand, and booking rate data within the preset review time window are determined, and the corresponding standard deviations of the price, demand, and booking rate data are obtained. Simultaneously, a preset scaling function is used to normalize the obtained standard deviations of the price, demand, and booking rate data to obtain the corresponding normalized standard deviations of the price, demand, and booking rate data (i.e., the volatility component of the price dimension), the normalized standard deviations of the demand, and the normalized standard deviations of the booking rate data (i.e., the volatility component of the booking rate dimension).
[0053] Furthermore, based on the obtained volatility components of the price dimension, demand dimension, and booking rate dimension, these components are combined to determine the corresponding environmental uncertainty measure, specifically:
[0054]
[0055] in: This is a measure of environmental uncertainty. For the volatility component in the price dimension, As a volatility component of the demand dimension, For the volatility component of the booking rate dimension, , , For component weights.
[0056] It is worth noting that the component weights in this embodiment can be specifically set according to actual needs, therefore, they are not specifically described in this embodiment. Also, the sum of the component weights in this embodiment is set to 1, i.e. .
[0057] In this embodiment, based on the decision state vector obtained in step S1.1 and the environmental uncertainty metric determined in step S1.2, the established value assessment model is dynamically adjusted, and corresponding candidate actions to be executed are determined. (Reference) Figures 3-5 This embodiment provides a dynamic generation method for exploration-exploitation strategies, which specifically includes the following steps:
[0058] Step S2.1: Model Construction. Based on the acquired historical hotel data, the daily market status, daily decision-making actions, and daily business revenue are determined. Simultaneously, the determined daily market status, daily decision-making actions, and daily business revenue are matched to obtain corresponding training data pairs. These training data pairs are then used to train the set Bayesian neural network to obtain the corresponding trained Bayesian neural network, which is the corresponding value assessment model.
[0059] Furthermore, the value assessment model in this embodiment includes a strategy configuration module to set the strategy parameters corresponding to the value assessment model. The strategy parameters are set according to the behavioral algorithm used by the value assessment model. Specifically, when the behavioral algorithm used by the value assessment model is an uncertainty-driven ε-greedy strategy, the corresponding strategy parameters include the base exploration probability and the uncertainty amplification coefficient. When the behavioral algorithm used by the value assessment model is a UCB strategy incorporating environmental uncertainty, the corresponding strategy parameters include the exploration weight coefficient.
[0060] Step S2.2: Action Evaluation. This involves determining the corresponding pricing and inventory control actions based on the hotel's revenue management operations (e.g., adjusting the publicly listed prices of different room types on different channels (e.g., large online travel agencies, official website), setting exclusive prices for different customer levels (e.g., members, contracted companies), dynamically adjusting the number of "reserved rooms," closing or opening sales channels during specific periods, and setting up consecutive stay discounts or package deals). Pricing actions include, but are not limited to, relative price adjustments (e.g., price increase of 5%, price decrease of 15%, price restoration to the base price), absolute pricing (i.e., setting an absolute fixed value), and competitive parameter pricing (e.g., setting the 25th percentile price of the competing group). Inventory control actions include, but are not limited to, quantity control (e.g., reserving 10 rooms for channel A and releasing 15 reserved rooms) and on / off control (e.g., closing sales of room type X for channel B and reopening sales of room type X for all channels). Simultaneously, based on the determined pricing and inventory control actions, a corresponding revenue management action list is created.
[0061] Furthermore, the decision state vector obtained in step S1.1.1 is paired one by one with each candidate action in the revenue management action list to construct corresponding vector action data pairs. Simultaneously, each constructed vector action data pair is used as input to the value assessment model to output the expected total revenue estimate and revenue uncertainty estimate corresponding to each vector action data pair.
[0062] Step S2.3: Decision Determination. Based on the strategy parameters set in Step S2.1, the corresponding exploration probability is set, and a uniform random number is generated using the pseudo-random number generator built into the value assessment model. Simultaneously, the corresponding execution logic is determined by comparing the exploration probability and the uniform random number. Based on the determined execution logic, the corresponding candidate actions to be executed are selected from the list of revenue management actions set in Step S2.2. Specifically, in this embodiment, the behavioral algorithm used by the value assessment model is an uncertainty-driven ε-greedy strategy, and its corresponding strategy parameters include the basic exploration probability and the uncertainty amplification coefficient. It is worth noting that the basic exploration probability and the uncertainty amplification coefficient in this embodiment can be specifically set according to actual needs; therefore, they are not specifically described in this embodiment.
[0063] Furthermore, based on the set basic exploration probability and the environmental uncertainty metric determined in step S1.2, the basic exploration probability and the environmental uncertainty metric are combined to obtain the corresponding exploration probability, specifically:
[0064]
[0065] in: To explore probability, This is a measure of environmental uncertainty. Based on the exploration of probability, This is an uncertainty amplification factor (which can be adjusted according to the hotel's risk appetite and the aggressiveness of its strategy, so it is not specifically described in this embodiment).
[0066] Furthermore, the pseudo-random number generator built into the value assessment model generates uniformly distributed random numbers within the interval [0,1). Simultaneously, the obtained exploration probability is compared with the generated uniformly distributed random numbers, and based on the comparison result, the corresponding execution logic is determined, specifically as follows:
[0067] When the exploration probability is greater than a uniform random number, the corresponding execution logic is the exploitation mode. At this time, based on all the expected total revenue estimates obtained in step S2.2, the maximum expected total revenue estimate is determined, and the candidate actions in the vector action data pair corresponding to the maximum expected total revenue estimate are set as the corresponding candidate actions to be executed.
[0068] Conversely, when the exploration probability is not greater than a uniform random number, the corresponding execution logic is exploration mode. In this case, based on all the revenue uncertainty estimates obtained in step S2.2, the median of the corresponding revenue uncertainty estimates is determined. All revenue uncertainty estimates are compared with the median, and revenue uncertainty estimates higher than the median are selected. Simultaneously, the selected revenue uncertainty estimates are combined to construct a corresponding high-uncertainty action candidate pool, and the sum of the corresponding candidate pool uncertainty estimates is obtained (i.e., combining all revenue uncertainty estimates in the high-uncertainty action candidate pool to obtain the corresponding sum of all revenue uncertainty estimates). Based on the ratio between each revenue uncertainty estimate in the high-uncertainty action candidate pool and the sum of the candidate pool uncertainty estimates, the selection probability corresponding to each revenue uncertainty estimate in the high-uncertainty action candidate pool is set. Simultaneously, based on the set selection probability, weighted random sampling is performed to obtain the corresponding candidate actions to be executed. That is, based on the revenue uncertainty estimates corresponding to the weighted random sampling results, corresponding vector action data pairs are determined, and the candidate actions in the determined vector action data pairs are set as the corresponding candidate actions to be executed.
[0069] In this embodiment, the strategy parameters set in step S2.1 can be combined with the environmental uncertainty metric determined in step S1.2 and the estimated revenue uncertainty obtained in step S2.2 to obtain the corresponding uncertainty potential value. The obtained uncertainty potential value is then combined with the estimated expected total revenue obtained in step S2.2 to obtain the corresponding comprehensive UCB score. Simultaneously, based on the comprehensive UCB score corresponding to each candidate action in the revenue management action list, the corresponding candidate action to be executed is determined. Specifically, the behavioral algorithm used by the value assessment model in this embodiment is a UCB strategy combined with environmental uncertainty, and its corresponding strategy parameters include an exploration weight coefficient. It is worth noting that the exploration weight coefficient in this embodiment can be specifically set according to actual needs, therefore, it is not specifically described in this embodiment.
[0070] Furthermore, based on the exploration weight coefficient set in step S2.1, the environmental uncertainty metric determined in step S1.2, and the estimated return uncertainty obtained in step S2.2, the exploration weight coefficient, the environmental uncertainty metric, and the estimated return uncertainty are combined to obtain the corresponding uncertainty potential value, specifically:
[0071]
[0072] in: This represents the potential value of uncertainty. To explore the weighting coefficients, This is a measure of environmental uncertainty. Valuation based on earnings uncertainty.
[0073] Furthermore, the obtained uncertainty potential value is combined with the expected total return estimate obtained in step S2.2 to obtain the sum of the uncertainty potential value and the expected total return estimate, which is the corresponding comprehensive UCB score. At the same time, the comprehensive UCB scores corresponding to each candidate action are compared to determine the highest comprehensive UCB score, and the candidate action corresponding to the highest comprehensive UCB score is set as the corresponding candidate action to be executed.
[0074] In this embodiment, the candidate actions to be executed determined in step S2.3 and the decision state vector obtained in step S1.1 are processed by a set demand forecasting model, and the corresponding inventory control instructions are determined by combining the set objective function. This embodiment provides a risk-constrained strategy execution method, which specifically includes the following steps:
[0075] Step S3.1: Model Prediction. This involves collecting relevant historical hotel data through the hotel's PMS, channel management platform, and market data sources. This data includes historical status (i.e., a snapshot of the market environment), historical actions (i.e., implemented pricing strategies), and historical results (i.e., actual bookings and cancellations). Simultaneously, based on the historical status, actions, and results of the hotel's historical data, corresponding training sample data pairs are constructed. These training sample data pairs are then used to train a probabilistic prediction model (such as a Bayesian neural network model) to obtain the corresponding trained probabilistic prediction model, which is the corresponding demand prediction model.
[0076] Furthermore, the decision state vector obtained in step S1.1 and the candidate actions to be executed determined in step S2.3 are both used as inputs to the demand forecasting model, so as to output the corresponding expected demand and probability distribution.
[0077] Step S3.2: Function Construction. This involves using the hotel's PMS (Property Management System) to obtain the real-time remaining salable inventory for each room type. Based on this inventory, the actual number of rooms sold for each type is determined. This actual number of rooms sold is then combined with the corresponding room price to determine the current immediate revenue value. Based on this current immediate revenue value, the candidate actions identified in Step S2.3, and the probability distribution obtained in Step S3.1, the corresponding objective function is set, specifically:
[0078]
[0079] in: Let X be the remaining inventory and t be the optimal value function when there are X decision stages remaining before the check-in date. For the candidate action to be executed, maximize the available operations (for example, if the candidate action to be executed is "only accept booking requests with a price not lower than P", then all the corresponding operable actions are to arbitrarily select one from the candidate price thresholds {P1, P2...Pn}). Let XD be the remaining inventory and the optimal value function corresponding to the decision stage t-1 days away from the check-in date. This represents the number of rooms sold in the current stage according to the probability distribution. This represents the current immediate reward value corresponding to the current decision stage t. For mathematical expectation operators, This is a risk aversion factor (which can be specifically set according to the hotel's aversion to revenue volatility risk, so it is not specifically described in this embodiment). This is the variance operator.
[0080] Furthermore, the objective function is optimized using a back-induction algorithm to obtain the corresponding optimal value function. Simultaneously, based on the remaining room inventory and remaining occupancy days corresponding to the optimal value function, these are combined with the expected demand and probability distribution obtained in step S3.1 to determine the corresponding final reserved inventory and current available quota. Furthermore, based on the candidate actions identified in step S2.3, the corresponding room pricing for these candidate actions is determined, and the final room pricing is set accordingly. In other words, based on the determined final reserved inventory, current available quota, and final room pricing, the corresponding inventory control instructions are constructed.
[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.
Claims
1. A hotel revenue management decision-making method based on machine learning, characterized in that, Including: S1: Quantification of uncertainty in environmental perception: Construct a decision state vector corresponding to the current market state through multi-source data from inside and outside the hotel, and determine the corresponding environmental uncertainty metric value based on the volatility of the decision state vector within a preset time window. S2: Dynamic generation of exploration-utilization strategy: Based on the decision state vector and the set value assessment model, obtain the expected total revenue valuation and revenue uncertainty valuation corresponding to each candidate action in the revenue management action list, and determine the corresponding candidate action to be executed through the environmental uncertainty metric. S3: Risk-constrained strategy execution: Based on the candidate actions to be executed and the decision state vector, determine the corresponding expected demand and probability distribution, construct the corresponding objective function, and obtain the corresponding inventory control instructions through the objective function.
2. The hotel revenue management decision-making method based on machine learning according to claim 1, characterized in that, The corresponding environmental uncertainty measures were determined, including: S1.1: Vector Construction: Collect and obtain corresponding internal and external multi-source data through internal systems, third-party data services and event calendars, and obtain corresponding pre-processed multi-source data through data cleaning and standardization. At the same time, determine the corresponding static features, dynamic market features, time series features and event features based on the pre-processed multi-source data, and concatenate the static features, dynamic market features, time series features and event features to obtain the corresponding decision state vector. S1.2: Value Determination: By collecting and acquiring all decision state vectors within a preset review time window, all corresponding dynamic market characteristics within the preset review time window are obtained, the corresponding dynamic market characteristic standard deviation is determined, and the corresponding normalized dynamic market characteristic standard deviation is obtained through a preset scaling function. At the same time, the corresponding environmental uncertainty metric is determined based on the normalized dynamic market characteristic standard deviation.
3. The hotel revenue management decision-making method based on machine learning according to claim 2, characterized in that, The dynamic market characteristics include price dimension data, demand dimension data, and booking rate dimension data. Based on the preset scaling function, the standard deviations of the price dimension data, demand dimension data, and booking rate dimension data are normalized to obtain the corresponding volatility components of the price dimension, demand dimension, and booking rate dimension. Simultaneously, these volatility components are combined to determine the corresponding environmental uncertainty measure, specifically: in: This is a measure of environmental uncertainty. For the volatility component in the price dimension, As a volatility component of the demand dimension, For the volatility component of the booking rate dimension, , , For component weights.
4. The hotel revenue management decision-making method based on machine learning according to claim 1, characterized in that, The corresponding candidate actions to be executed have been identified, including: S2.1: Model building: By constructing training data pairs based on the daily market status, daily decision-making actions, and daily business revenue generated from the hotel's historical data, a Bayesian neural network is trained using the training data pairs to obtain the corresponding value assessment model. S2.2: Action Evaluation: Based on the hotel's corresponding revenue management operations, determine the corresponding pricing actions and inventory control actions, set up the corresponding revenue management action list, and pair the decision state vector with each candidate action in the revenue management action list one by one to construct the corresponding vector action data pair. The vector action data pair is used as the input of the value evaluation model, and the output is to obtain the corresponding expected total revenue valuation and revenue uncertainty valuation. S2.3: Decision Determination: Based on the strategy parameters set according to the value assessment model, the expected total revenue valuation and revenue uncertainty valuation are processed, and based on the processed expected total revenue valuation and revenue uncertainty valuation, the corresponding candidate actions to be executed are determined from the revenue management action list.
5. The hotel revenue management decision-making method based on machine learning according to claim 4, characterized in that, The value assessment model includes a strategy configuration module, which sets corresponding strategy parameters based on the behavioral algorithm used by the value assessment model. Specifically: When the behavior algorithm is an uncertainty-driven ε-greedy policy, the corresponding policy parameters include the base exploration probability and the uncertainty amplification coefficient; when the behavior algorithm is a UCB policy that incorporates environmental uncertainty, the corresponding policy parameters include the exploration weight coefficient.
6. The hotel revenue management decision-making method based on machine learning according to claim 5, characterized in that, When the strategy parameters include a base exploration probability and an uncertainty amplification coefficient, the base exploration probability and the environmental uncertainty metric are combined to obtain the corresponding exploration probability. A corresponding uniform random number is then set using the pseudo-random number generator built into the value assessment model. Simultaneously, based on the comparison between the exploration probability and the uniform random number, the corresponding execution logic is determined, specifically: When the exploration probability is greater than a uniform random number, the corresponding execution logic is the exploitation mode; otherwise, the corresponding execution logic is the exploration mode.
7. The hotel revenue management decision-making method based on machine learning according to claim 6, characterized in that, When the execution logic is in the utilization mode, the expected total revenue valuation corresponding to each candidate action is compared to determine the maximum expected total revenue valuation, and the candidate action corresponding to the maximum expected total revenue valuation is set as the candidate action to be executed. When the execution logic is in exploration mode, the median of the revenue uncertainty estimate is determined based on the revenue uncertainty estimate corresponding to each candidate action. A high-uncertainty action candidate pool is constructed using revenue uncertainty estimates higher than the median revenue uncertainty estimate. The sum of the uncertainty estimates of the candidate pool is obtained. Simultaneously, the selection probability corresponding to each revenue uncertainty estimate in the high-uncertainty action candidate pool is set based on the ratio between the revenue uncertainty estimate and the sum of the uncertainty estimates of the candidate pool. Weighted random sampling is performed based on the selection probability, and the candidate action corresponding to the revenue uncertainty estimate of the weighted random sampling result is set as the corresponding candidate action to be executed.
8. The hotel revenue management decision-making method based on machine learning according to claim 5, characterized in that, When the strategy parameters include an exploration weight coefficient, the exploration weight coefficient, the environmental uncertainty metric, and the return uncertainty estimate are combined to obtain the corresponding uncertainty potential value. The uncertainty potential value is then combined with the expected total return estimate to obtain the corresponding comprehensive UCB score. Simultaneously, the comprehensive UCB scores corresponding to each candidate action are compared to determine the highest comprehensive UCB score. The candidate action corresponding to the highest comprehensive UCB score is then set as the corresponding candidate action to be executed.
9. The hotel revenue management decision-making method based on machine learning according to claim 1, characterized in that, Obtain the corresponding inventory control instructions, including: S3.1: Model prediction: By constructing training sample data pairs based on the historical states, actions, and results corresponding to the hotel's historical data, a probabilistic prediction model is trained using the training sample data pairs to obtain a demand prediction model. At the same time, the decision state vector and the candidate actions to be executed are both used as inputs to the demand prediction model, and the corresponding expected demand and probability distribution are output. S3.2: Function Construction: Based on the actual number of rooms sold and the room price for each room type in the hotel, determine the corresponding current immediate revenue value. Combine the current immediate revenue value, expected demand, and probability distribution to set the corresponding objective function. Simultaneously, optimize the objective function using a back induction algorithm to obtain the optimal value function. Combine the remaining room inventory and remaining stay days corresponding to the optimal value function with the expected demand and probability distribution to determine the corresponding final reserved inventory and current available quota. At the same time, construct the corresponding inventory control instruction based on the room price, final reserved inventory, and current available quota corresponding to the candidate action to be executed.
10. A hotel revenue management decision-making system based on machine learning, characterized in that, The hotel revenue management decision-making method based on machine learning, as described in any one of claims 1-9, was used.