CN111445054A - Evacuation path recommendation method based on machine learning - Google Patents

Evacuation path recommendation method based on machine learning Download PDF

Info

Publication number
CN111445054A
CN111445054A CN201911072786.7A CN201911072786A CN111445054A CN 111445054 A CN111445054 A CN 111445054A CN 201911072786 A CN201911072786 A CN 201911072786A CN 111445054 A CN111445054 A CN 111445054A
Authority
CN
China
Prior art keywords
data
machine learning
method based
reduction
recommendation method
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911072786.7A
Other languages
Chinese (zh)
Inventor
黄元琪
毕重科
侯敏
王佳敏
傅宝锋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianjin University
Original Assignee
Tianjin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianjin University filed Critical Tianjin University
Priority to CN201911072786.7A priority Critical patent/CN111445054A/en
Publication of CN111445054A publication Critical patent/CN111445054A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G06Q10/047—Optimisation of routes or paths, e.g. travelling salesman problem
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G06Q50/265—Personal security, identity or safety

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Development Economics (AREA)
  • Molecular Biology (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Game Theory and Decision Science (AREA)
  • Mathematical Analysis (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Computational Mathematics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Algebra (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Security & Cryptography (AREA)
  • Educational Administration (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an evacuation path recommendation method based on machine learning, which comprises the following steps: step one, acquiring weights of different factors in data by using incremental training of an automatic encoder; secondly, distributing weights of different dimensions by using a back propagation optimization data reduction process, and keeping important features in low-dimensional data; step three, visualizing the data by using a 2D scatter diagram after the data reduction is finished, and evaluating the performance together with other dimension reduction methods; designing an evacuation path recommending Markov reward function and a discount factor; designing and realizing refuge path prediction and carrying out visualization; the invention enlarges the range of Markov process application and path design through the cooperation of data reduction and machine learning, so that high-dimensional data from different sources can be subjected to path prediction after processing, and meanwhile, the incremental self-encoder model used simplifies input data and improves the efficiency and precision of the machine learning process.

Description

Evacuation path recommendation method based on machine learning
Technical Field
The invention relates to the technical field of earthquake refuge path methods, in particular to an earthquake refuge path recommendation method based on machine learning.
Background
The evacuation route has important significance for large-scale emergency safety management. Under the guidance of a good evacuation route, people can save many lives and properties, particularly in disasters where people cannot find a way intuitively, such as dense smoke, earthquakes, nuclear pollution and the like. A good evacuation route solution should be a globally optimal solution that can be quickly obtained in a disaster environment. However, there are a considerable number of factors to consider in a disaster. If each factor is one-dimensional to the disaster data, then the disaster data is high-dimensional. When the refuge path is involved, a plurality of data indexes need to be considered, for example, in an earthquake, the earthquake series can not be considered as the only index, a plurality of indexes are combined together to obtain a result, and the common action of a high-dimensional data processing technology and a machine learning method is needed.
There have been several disaster big data evacuation route design works, and especially with recent years, big data of extreme events has become an important topic, and a great deal of research has been conducted, which can provide help for emergency evacuation management. The most effective way to help us escape from a hazardous area is to design an optimal evacuation route immediately upon the occurrence of a disaster. However, there are many factors that affect a disaster. Analysis of such high-dimensional complex data is difficult to directly analyze. Many researchers have therefore proposed optimal evacuation routes in different situations. For example, evacuation schemes for places with dense pedestrian flows, such as stadiums, hospitals, airports, stations and the like, when a fire occurs, mainly take the pedestrian flow and the risk of pressing and treading into consideration. There are also methods for traffic evacuation assistance via smart devices, information communication, and the study of dynamic crowd evacuation cases. For natural disasters, there are also some presence-based studies, targeting a fixed disaster with sufficient characteristics, such as typhoon. However, the above methods can be used only in special cases, that is, in cases where the data is not so complex and there is some a priori rule to be followed. The present invention thus contemplates a general evacuation route recommendation method in earthquake disasters. For this reason, large data must be processed quickly, and data reduction is one of the most efficient methods, especially when global data is considered, as is the case with optimal evacuation route design. The data reduction method for emergency management should fully retain important features and take the weight of all factors into consideration. Some data compression methods, which are typically used for image processing or real-time compression, can greatly reduce the amount of data but cause feature loss. For example, fpzip is an effective way of compressing float. The data size of the image can be greatly reduced, but the actual error caused by compression is too large to be used for emergency management. There are also some dimension reduction methods, which are widely used in the field of visualization. The most common linear dimensionality reduction methods are, for example, principal component analysis, tensor decomposition, and proper orthogonal decomposition. These methods have the advantage of efficiency and the disadvantage of large errors. Some non-linear methods, although with small errors, do not meet the emergency requirements in terms of efficiency. For example, a dynamic pattern decomposition method isomap for flow data or time series data can reduce the dimension of complex data, MDS can be used to obtain the relative position between two sampling points, and TSNE designs a T-distribution to match the characteristics of high-dimensional data. These data dimension reduction methods have difficulty in satisfying both accuracy and efficiency requirements. The invention provides a data dimension reduction method based on an automatic encoder, which can keep the important characteristics of high-dimensional data and has higher efficiency. And based on the dimension reduction data, an optimal evacuation path prediction method based on a Markov decision process is provided.
Disclosure of Invention
A quick earthquake shelter evacuation route recommendation method framework is provided. The main idea of the method comprises two parts: a feature retention data dimension reduction method and a global optimal evacuation path prediction method. Figure 1 shows the workflow of the method of the invention. Firstly, the high-dimensional big data is reduced by using a characteristic data dimension reduction method, which is realized by using an automatic encoder based on multi-layer perception. By using a two-dimensional scatter diagram to visualize the data after dimensionality reduction, we can obtain the performance comparison of the method with other data dimensionality reduction methods. In the dimension reduction process, the problems of gradient explosion and gradient disappearance are solved by utilizing a SoftPlus function and an Xavier. Meanwhile, an incremental model is provided to meet the requirement of rapid model evolution. Then, in a method for predicting the recommended evacuation route by using a Markov process, a corresponding reward function and a discount factor are designed to achieve global optimization of the evacuation route.
The technical scheme of the invention is as follows:
a machine learning-based earthquake refuge path recommendation method comprises the following steps:
firstly, obtaining weight coefficients of different factors from meteorological data by using an automatic encoder in cooperation with an incremental training model;
distributing weights of different dimensions by using a back propagation optimization data reduction process, and keeping important features in low-dimensional data;
evaluating the data processed by the method and the data dimension reduction method by using a visualization tool;
step four, designing a Reward function Reward (s, a) and a discount factor (gamma) of the Markov process;
and fifthly, analyzing the data to obtain a conclusion and performing visual processing on an analysis result.
The back propagation gradient optimization method used in the second step makes the data conform to the distribution:
Figure RE-GDA0002391922130000021
the gradient problem was optimized using the activation function SoftPlus:
σ(x)＝log(1+ex)。
so that the training data for each fully connected layer can be obtained from the previous layer:
σX[i]＝W*Zi-1+b
Z[j]＝σ(X[j])。
the visualization tool in the third step evaluates the data processed by the method and other data dimension reduction methods, and the conclusion is shown in fig. 4; the design of Reward function Reward (s, a) and discount factor (γ) in the markov process in the fourth step is as follows:
Figure RE-GDA0002391922130000031
and the results of the visualization processing of the refuge path results in the fifth step are shown in fig. 9.
Advantageous effects
Compared with the prior art, the invention has the advantages that:
a method for rapidly detecting an optimal evacuation route from high-dimensional, complex and big data is provided. The advantages of the method are mainly focused on two parts: feature preserving data reduction and optimal path detection. Aiming at the characteristic retention data dimension reduction part, the invention provides a data dimension reduction method based on a multilayer automatic encoder. The method can greatly reduce the high-dimensional and complex data volume. At the same time, our autoencoder approach retains important characteristics. By maintaining the data distribution during initialization, the use of the SoftPlus activation function with Xavier in the optimization process solves the gradient extinction and explosion problems. In addition, the data reduction method is also highly efficient. Because the incremental training model is used, the data can be processed only by training new data, so that the method can achieve higher dimension reduction compression efficiency, and finally, the simplified data is visualized by using the two-dimensional scatter diagram, which shows that the dimension reduction method can fully reserve the important characteristics of high-dimensional and complex big data. For the optimal path detection part, the invention provides a Markov decision process-based method, and a new action rule, a reward function and a discount factor are designed for the evacuation route prediction of meteorological data. In addition, the invention performs visual analysis on all possible paths and provides the best path in consideration of the probability distribution problem.
Drawings
Fig. 1 is a flowchart of an evacuation path recommendation method based on machine learning according to the present invention.
Fig. 2 is a flow chart of an autoencoder used in the present invention.
FIG. 3 is a schematic diagram of a portion of a meteorological data distribution interval used in the present invention.
FIG. 4 is a two-dimensional scatter diagram of data distribution after being processed by the auto-encoder used in the present invention and the existing data dimension reduction method.
FIG. 5 is a graph of the accuracy of classifying data by different machine learning methods after the data is processed by the automatic encoder used in the present invention and the existing data dimension reduction method.
FIG. 6 is a diagram of a Markov decision used in the present invention.
FIG. 7 is a diagram illustrating the effect of different discount factors on the experimental results.
Fig. 8 shows a simplified result of data visualization using thermodynamic diagrams in the present invention, and the data is meteorological data of 5 provinces around tokyo in japan from 9/3/2011 to 10/3/s (311 am). Here, the starting point of (a) and (b) is Tokyo; (c) and (d) is Nanbu; (e) and (f) is Kitaibaraki. And (3) reducing the dimension of the data of the high-dimensional data in 3 months and 9 days. (a) The results of (c), (e) and (c) can be considered as true because they are the results of training the model using 3 months and 9 days of data and bringing in the same day of data. On the other hand, (b), (d) and (f) use data from 3 months and 10 days to bring in the training model for the previous day's data. It can be seen that the results of (b), (d) and (f) are almost the same as those of (a), (c) and (e). Thus, the model trained using our method can be used to reduce other datasets.
FIG. 9 is a visualization of the recommended route results in the present invention, starting points are (a) chichichibu, (b) Shimotsuma, and (c) Kitaibaraki, respectively, and the starting points are marked with red arrows. For each starting point, several optimal routes are recommended according to different requirements.
Detailed Description
The invention is further illustrated by the following specific examples and the accompanying drawings. The examples are intended to better enable those skilled in the art to better understand the present invention and are not intended to limit the present invention in any way.
As shown in FIG. 1, the invention provides earthquake evacuation path recommendation based on machine learning, which comprises the following steps:
step one 101, using an automatic encoder to carry out incremental training to obtain the weight of each factor in data;
for natural disasters, the similarity between two cities is the most important feature. This means that the reduced dimension data should maintain the similarity, which is determined by all the factors and their weights, the weight of each factor being very important. For example, if the magnitude and temperature of an earthquake rise from 5 to 6, the former can have catastrophic consequences, while the latter is negligible in our daily lives. In order to solve the problem, the invention provides a data compression automatic encoder method. The data dimensionality reduction comprises two steps, namely: performing data training by using an automatic encoder; II, secondly: data simplification and visualization.
Fig. 2 is a workflow of automatic encoding. Through the learning of multi-layer perception, the weights of different factors in the meteorological data can be automatically obtained. As shown, high dimensional data is input as source data, which is converted to vectors. The dimensionality of the vector will be reduced layer by layer through multi-layer perception. Wherein each layer is a fully connected layer:
Figure RE-GDA0002391922130000041
after the dimension is reduced to two dimensions and visualization can be carried out through the scatter diagram, the obtained low-dimensional data is reconstructed again by adopting the method the same as the dimension reduction process. Then, comparing the reconstructed data with the source data, and calculating the cross entropy loss:
Figure RE-GDA0002391922130000051
finally, the data reduction process is optimized using back propagation. In the optimization process, weights for different dimensions in the source data will be assigned. This means that the important features will remain entirely in the low dimensional data. The present invention processes big data using an incremental training process. Training large amounts of data together can lead to two important problems, one: computing big data is very inefficient; II, secondly: overfitting; both of these problems can be solved by an incremental process. Meanwhile, in order to improve the robustness of the self-coding method, Gaussian noise is added in the training process.
Step two 102, solving the problem of gradient disappearance or gradient explosion in the self-encoder training through the data reduction optimization process;
when the method is applied to meteorological data, the problems of gradient explosion and gradient disappearance appear. For example, the weight of the air pressure assignment is large, affecting the output. Then, the gradient vanishing problem occurs during the back propagation process, and the optimization process is difficult to achieve convergence. The present invention addresses both of these issues, with each layer in the initialization process and the optimization process. During initialization, our goal is to have the input and output distributions of each layer nearly identical. Here, our method uses the Xavier method, which preserves the distribution of the inputs: for example, a Gaussian distribution in the meteorological data, such as plot three, should be retained in the output. Firstly, different meteorological elements are normalized by preprocessing. Here, a standard scaler approach is used. Let the parameters of the fully connected layer be W and b, where W is the weight matrix and b is the bias vector. To maintain the distribution characteristics, W should satisfy a uniform distribution as shown in the equation:
Figure RE-GDA0002391922130000052
wherein U represents a uniform distribution. Further, n isinAnd noutThe dimensions of the input and output, respectively. In order to solve the gradient problem in the optimization process, the gradient of the activation function only needs to be about 1, because the reason for the disappearance of the gradient is that the gradient of the activation function is too small. Also, gradient explosions are caused by activation functions of large gradients. Thus, the present invention uses the SoftPlus function as the activation function because when x is>At 0, its gradient σ (x) is almostEqual to 1, as follows:
σ(x)＝log(1+ex)
finally, the training data for each fully-connected layer may be obtained from the previous layer output result equation:
σX[i]＝W*Zi-1+b
Z[j]＝σ(X[j]).
step three 103, evaluating the data processed by the method and other data dimension reduction methods by using a visualization tool; this section includes two aspects, as shown in FIG. four, we will use a two-dimensional scatter plot to visualize the data reduction results. The figure shows one reduction result. The data set is meteorological data for 57 cities around tokyo at 3/10 (311 geodetic epicranium) 2011. Fig. 4(a) is a visualization of the original high dimensional dataset with colors representing weather similarities between cities and Nerima. The legend colors range from red to blue, with red meaning that the weather between two cities is the same and blue representing a difference. The visualization results of the data after dimensionality reduction are shown in FIG. 4(b) the method employed in the present invention, (c) PCA, (d) MDA, (e) Isomap, and (f) TSNE. It can be seen from the results that the method employed by the present invention most closely resembles the original high dimensional data. The resulting similarities between cities may be used for emergency management evacuation route prediction.
Meanwhile, in order to better illustrate the advantages of the method in data compression, quantitative research is carried out on classification accuracy of the method after data compression by using some data sets, and the research result is shown in fig. 5.
Step four 104, designing Reward functions Reward (s, a) and discount factors (gamma) of a Markov process, wherein a Markov Decision Process (MDP) method can provide a global optimal path for emergency management, particularly natural disaster emergency management, and the core idea of the method is shown in figure six; the MDP is a 5-tuple (S, A, P, R, γ). Herein, as an input, S is a set of cities; a is the set of currently possible evacuation actionsCombining; p is the probability of using one of the actions (a) to transition from one city to another. In the invention, we set up
Figure RE-GDA0002391922130000061
Since we determine that the destination can be reached. R is the reward function. It is the most important factor for MDP. In the present invention, r refers to a reward for a person to go to another city. γ is a discount factor, which is added to be used to solve the global optimal path. For these 5 parameters, it can be seen from fig. six that MDP is a dynamic programming process that acquires v (S) and pi (S). π(s) refers to the current city s, and v(s) represents the jackpot for the next determined city s.
Action (a) designed for evacuation route prediction requires the definition of a state transition rule, which is influenced by real life. In the present invention we assume that state transitions only occur between two adjacent cities. For example, city s1 has k neighboring cities, then there are only k options in the next step. When going to other cities, one of the cities must be selected as a stopover.
The reward function R designed for evacuation route prediction in the invention should consider the following two important factors: the environment after transfer should become better; the time and distance cost of transfer should be as small as possible; the invention designs a reward function according to the two factors:
Figure RE-GDA0002391922130000071
where dif (s, s ') is the variance of the risk score between cities s and s'. As mentioned above, the risk score of a certain city is obtained by data dimension reduction by using a self-coding method. If the risk score becomes lower after the transition, the prize value will support the transition as the value of dif (s, s') becomes larger. dis (s, s') is the Euclidean distance between two cities. The longer the distance, the smaller the prize value.
Next, the evacuation route prediction discount factor γ of the present invention will be described. The final decision will be affected by the discount factor, as shown in the equation:
Figure RE-GDA0002391922130000072
Figure RE-GDA0002391922130000073
the decision in step t is the sum of the rewards in the current step and the next several steps, the weight of which is determined by the discount factor. Thus, when the discount factor is larger, the number of predicted routing steps will be larger. Fig. 7(a) and (b) predict two evacuation routes with the starting point set to tokyo, the difference between (a) and (b) being due to their discount factors, which are 0.1 and 0.9, respectively. Obviously, when γ is 0.9, more steps are selected. This is also an interface that the user can select to set the discount factor smaller if the person wants to select a safe evacuation location nearby. Similarly, when a person wants to go to a far away safe city, the discount factor may be set larger.
Using the action rules a, reward function R and discount factor γ we have designed, we can use MDP to recommend the best evacuation route. Finally, the invention proposes a probability distribution function to display possible subsequent steps of different user needs, namely:
Figure RE-GDA0002391922130000074
for example, some people want to find a safe evacuation location near the current city, while some people want to go to a city that is as safe as possible. The overlapping parts of the different routes are the most important evacuation routes. Several such overlapping routes are shown next.
Step five 105, the present invention experimented with demonstrating the effectiveness of the evacuation route advisory framework by using the meteorological data in japan. We will describe the results of the data dimension reduction method and the route recommendation method, respectively.
To demonstrate that the training model of the present invention can effectively reduce data volume, we performed two experiments (e1, e2) using two data sets (a, B). Wherein the data A is meteorological data of 3 months and 9 days in 2011; data B is meteorological data of 3 months and 10 days 2011; e1, using data A as training data, and performing dimensionality reduction on A; e2 is to use data A as training data and then perform dimension reduction on B. Thus, the result of e1 can be considered as a ground truth. We can compare the results of e2 and e 1. The visualization results are shown in fig. 8. From 3 sets of visualization results (different starting points), we can see that our training model can be fully used for dimensionality reduction of other datasets. Therefore, our data dimension reduction method can successfully retain important characteristics into the dimension-reduced dataset.
Next, the feasibility of visualizing the calculated optimal path using the markov decision process method is explained. The data used was meteorological data for 3 months and 10 days 2011; three sets of experiments were performed by selecting three different starting points: chichichibu, Shimotsuma and Kitaibaraki. For each set of data, we provide three best paths according to different requirements, as shown in fig. 9. The obtained path is compared with the original data to illustrate the effectiveness of the refuge recommended path, and the following table shows that:
Table2:The parameters in part of one evacuation route,which startsfrom Chichibu
Figure RE-GDA0002391922130000081
the top to bottom data in the table is the direction of the evacuation route and these raw data may indicate that all locations on the route are safe. For example, precipitation is 0 in all locations, as rain can spread nuclear pollution there. Meanwhile, the air temperature on the route is also relatively appropriate.
The most important advantage of the evacuation route recommendation of the present invention is the ability to provide multiple safe routes in time during a disaster. Meanwhile, inputting high dimensional data means that most possible parameters are fully considered in the safe route recommendation process. Meanwhile, through the incremental self-encoder model, the invention can meet the requirement of time efficiency while ensuring the precision.
It should be understood that the embodiments and examples discussed herein are illustrative only and that various modifications or changes in light thereof will be suggested to persons skilled in the art and are to be included within the spirit and purview of this application and scope of the appended claims.

Claims (4)

1. An evacuation path recommendation method based on machine learning is characterized by comprising the following steps:
firstly, obtaining weight coefficients of different factors from meteorological data by using an automatic encoder in cooperation with an incremental training model;
distributing weights of different dimensions by using a back propagation optimization data reduction process, and keeping important features in low-dimensional data;
step three, after the data reduction is completed, visualizing the data by using a 2D (two-dimensional) scatter diagram, and evaluating the performance together with a dimension reduction method;
step four, designing a Reward function Reward (s, a) and a discount factor (gamma) of the Markov process;
and fifthly, analyzing the data to obtain a conclusion and performing visual processing on an analysis result.
2. The evacuation path recommendation method based on machine learning according to claim 1, wherein the back propagation gradient optimization process used in the second step conforms the data to the distribution:
Figure RE-FDA0002514125160000011
the gradient problem was optimized using the activation function SoftPlus:
σ(x)＝log(1+ex)
so that the training data for each fully connected layer can be obtained from the previous layer:
σX[i]＝W*Zi-1+b
Z[j]＝σ(X[j])。
3. the evacuation path recommendation method based on machine learning as claimed in claim 1, wherein the step three is visualization and comparison of dimension reduction results of the auto-encoder and other data in the method of the present invention.
4. An evacuation path recommendation method based on machine learning as claimed in claim 1, wherein the Reward function Reward (s, a) and the discount factor (γ) are designed in Markov process in the fourth step, that is:
Figure RE-FDA0002514125160000012
CN201911072786.7A 2019-11-05 2019-11-05 Evacuation path recommendation method based on machine learning Pending CN111445054A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911072786.7A CN111445054A (en) 2019-11-05 2019-11-05 Evacuation path recommendation method based on machine learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911072786.7A CN111445054A (en) 2019-11-05 2019-11-05 Evacuation path recommendation method based on machine learning

Publications (1)

Publication Number Publication Date
CN111445054A true CN111445054A (en) 2020-07-24

Family

ID=71650613

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911072786.7A Pending CN111445054A (en) 2019-11-05 2019-11-05 Evacuation path recommendation method based on machine learning

Country Status (1)

Country Link
CN (1) CN111445054A (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101868811A (en) * 2007-09-19 2010-10-20 联合工艺公司 System and method for threat propagation estimation
CN110348969A (en) * 2019-07-16 2019-10-18 哈尔滨工程大学 Taxi based on deep learning and big data analysis seeks objective policy recommendation method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101868811A (en) * 2007-09-19 2010-10-20 联合工艺公司 System and method for threat propagation estimation
CN110348969A (en) * 2019-07-16 2019-10-18 哈尔滨工程大学 Taxi based on deep learning and big data analysis seeks objective policy recommendation method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
CHONGKE BI ET AL.: ""Evacuation route recommendation using auto-encoder and Markov decision process"" *

Similar Documents

Publication Publication Date Title
Wu et al. Examining the sensitivity of spatial scale in cellular automata Markov chain simulation of land use change
Naik et al. Computer vision uncovers predictors of physical urban change
Yu et al. Prediction of highway tunnel pavement performance based on digital twin and multiple time series stacking
Long et al. Mapping block-level urban areas for all Chinese cities
Ebrahimian Ghajari et al. Spatial modelling of urban physical vulnerability to explosion hazards using GIS and fuzzy MCDA
Zhao et al. Recognition of building group patterns using graph convolutional network
Son et al. Developing scenario-based technology roadmap in the big data era: an utilisation of fuzzy cognitive map and text mining techniques
Jiang et al. Large-scale taxi O/D visual analytics for understanding metropolitan human movement patterns
Shen et al. Understanding the spatial–temporal variation of human footprint in Jiangsu Province, China, its anthropogenic and natural drivers and potential implications
Huang et al. Research on urban modern architectural art based on artificial intelligence and GIS image recognition system
Ding et al. Hyperautomation for air quality evaluations: A perspective of evidential three-way decision-making
Ahmed et al. Learning to extract buildings from ultra-high-resolution drone images and noisy labels
Kulagin et al. Designing an educational intelligent system with natural language processing based on fuzzy logic
Poon Quantitative methods: past and present
CN117217602A (en) Generalized and automatic unmanned aerial vehicle task efficiency evaluation index construction method
Kontokosta et al. Energyviz: Web-based eco-visualization of urban energy use from building benchmarking data
Olszewski et al. Using fuzzy geoparticipation methods to optimize the spatial development process in a smart city
CN111445054A (en) Evacuation path recommendation method based on machine learning
Al_Sayed et al. On the nature of urban dependencies: How Manhattan and Barcelona reinforced a natural organisation despite planning intentionality
Wang et al. Pyramid self-attention mechanism-based change detection in hyperspectral imagery
CN114819140A (en) Model pruning method and device and computer equipment
Ishikawa et al. Towards construction of an explanation framework for whole processes of data analysis applications: concepts and use cases
Jin Multiple solutions of the Kirchhoff-type problem in
Du et al. Mining multicity urban data for sustainable population relocation
Ishikawa et al. Generalized difference method for generating integrated hypotheses in social big data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200724

WD01 Invention patent application deemed withdrawn after publication