A human-like cognitive reasoning framework for user demand prediction in a pension context

Abstract

The application discloses a human-like cognitive reasoning framework for user demand prediction in a pension situation, belongs to the field of information technology, artificial intelligence and man-machine interaction, and discloses the human-like cognitive reasoning framework for user demand prediction in the pension situation, which comprises a resource storage, an encoder, a decoder, an evaluation updater, an intuition generator, a cognitive device, a training updater and a decision maker. The application provides the human-like cognitive reasoning framework for user demand prediction in the pension situation, which can be applied in the pension situation, enables the service robot to have accurate and efficient reasoning capability, and can predict the specific demand of the old people under the premise that limited external conditions are provided, so that the humanization of the service robot is of great significance and has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the fields of information technology, artificial intelligence, and human-computer interaction, and specifically relates to a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios based on human reasoning mechanisms. Background Technology

[0002] Robots lack the reasoning abilities of humans, leading to poor human-computer interaction experiences. Current methods for reasoning about user needs mostly rely on training fixed models with fixed data, failing to achieve "cognitive expansion" and only addressing fixed types of needs. They cannot adapt to evolving needs arising from user preferences. Therefore, proposing a human-like cognitive reasoning framework that can continuously adapt to evolving needs based on user preferences is both necessary and urgent. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to propose a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios.

[0004] The technical solution of this invention is implemented as follows: a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios, comprising:

[0005] Resource storage: Used to store reliable data pairs, record established facts and predict successful cases, and also for counting and accessing cases and objects;

[0006] Encoder: Used to encode plain text objects, transforming plain text into a matrix of word vectors that can be computed by neural networks;

[0007] Decoder: Used for decoding the prediction results of neural networks, decoding the prediction results into a list consisting of the text format corresponding to the object with the highest non-linear probability and its non-linear probability;

[0008] Evaluation Updater: Used to integrate the input and the framework's output for evaluation. There are two evaluation methods: one is manual evaluation, where the framework interacts with the user to inquire about the user's actual needs; the other is data-driven evaluation, which is based entirely on the results of historical data. If the evaluation result is positive, the data is stored in the resource storage; otherwise, the resource storage is traversed, and if the same data is found, the data is deleted or the data obtained from inquiring with the user is stored, and the time index of the resource storage is updated.

[0009] Intuition Generator: Composed of a habit model and a short-term memory model, it is used to simulate human intuition. The output of both is a list of prediction results and nonlinear probabilities. The habit model assigns a nonlinear probability to the prediction result based on the distribution of the most frequent prediction result in the resource memory. The short-term memory model assigns a nonlinear probability to the prediction result based on the distance of the corresponding prediction result in the resource memory on the time axis. If there is no corresponding prediction result for the input in the resource memory, it means that this framework has no memory of this output, that is, the short-term memory model has no output.

[0010] Cognitive device: Composed of a cognitive model and a prior model, it is used to simulate the human cognitive process. It expands or reduces the output nodes of the cognitive model according to the number of object types in the resource storage. It is also used to identify the predicted value corresponding to the data output of the encoder. The predicted value is processed by the decoder to obtain the corresponding output. The cognitive device also has a special mechanism that when the number of nodes of the cognitive model changes, the weights and structure of the trained cognitive model are overlaid on the prior model.

[0011] Training updater: The built-in encoder is used to package the data in the resource storage, encode it, and then use it to train and update the network weights of the cognitive model in the cognitive device.

[0012] Decision Maker: It aggregates the outputs of all modules and selects the object with the highest non-linear probability as the final output.

[0013] As a preferred embodiment, the data pair described in this invention is "input-prediction result", and the access time-indexed records have index numbers ranging from 0 to the maximum.

[0014] As another preferred embodiment, the word vectors described in this invention are encoded using "-1", "0", and "1".

[0015] As another preferred embodiment, if the encoded text object is insufficient to fill the entire 32*32 matrix, the encoder of the present invention uses a method of evenly partitioning the matrix to copy the encoded data to fill it.

[0016] As another preferred embodiment, the decoder of the present invention predicts results as a dictionary consisting of category numbers and corresponding results, and a nonlinear probability corresponding to each number.

[0017] As another preferred embodiment, the evaluation updater of the present invention outputs a prediction result, and a positive evaluation result means that the user is satisfied or the data is correct.

[0018] Secondly, in the cognitive device described in this invention, the prior model refers to a neural network trained with prior knowledge, the cognitive model refers to a neural network with a Softmax layer that can output nonlinear probabilities, and the output refers to a list composed of the output and the nonlinear probabilities.

[0019] In addition, in the training updater described in this invention, data refers to input-prediction results, and cognitive model refers to neural networks.

[0020] The method for predicting a natural person's demand for fruit using the above framework includes the following steps:

[0021] Step 1): Train the cognitive model in the cognitive device using the training updater based on the historical data set in the resource store, and load the weights of the prior model in the cognitive device;

[0022] Step 2): Use an encoder to transcode the input plain text external conditions, and then input the transcoded data into the cognitive device;

[0023] Step 3): Input the outputs of the cognitive model and the prior model in the cognitive device into the decoder to obtain two lists consisting of the prediction results and nonlinear probabilities representing their occurrence probability;

[0024] Step 4): Based on the distribution of historical data in the resource storage, find the prediction result with the highest frequency and assign it a non-linear probability P1 as the output of the habit model in the intuition generator. If there is a corresponding prediction result in the input resource storage, assign that result as the output of the short-term memory model in the intuition generator and assign it a non-linear probability P2.

[0025] Step 5): Integrate the outputs of the cognizer and the intuition generator into the decision-maker to obtain the final prediction result;

[0026] Step 6): Input the input and the final prediction result into the evaluation updater. If the evaluation result is positive, proceed to step 7); otherwise, proceed to step 8).

[0027] Step 7): Increment the maximum time index number of the resource manager by one, and store the input and the final prediction result at the maximum time index of the resource storage;

[0028] Step 8): Search the resource manager. If a data set consisting of the input and the final prediction result exists in the resource manager, delete the data and update the time index number.

[0029] Step 9): If a new object is input, repeat steps 1 through 8.

[0030] Advantages of this invention: Current reasoning algorithms used in service robots lack human-like reasoning capabilities and cannot adapt to new output objects based on user preferences, potentially leading to a poor human-computer interaction experience. This invention proposes a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios. This framework can change its model structure to adapt to changing needs, achieving a human-like cognitive mechanism. Furthermore, it considers the influence of human psychological characteristics on choices and uses an intuition generator to simulate human intuition, which is of great significance for service robot service optimization and has broad application prospects. Attached Figure Description

[0031] Figure 1 This is a structural block diagram of a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios, as proposed in this invention.

[0032] Figure 2 This is a flowchart of a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios, as proposed in this invention.

[0033] Figure 3 This is a practical result of the present invention, which is a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios.

[0034] Figure 4 This is the pseudocode for the gradient descent principle of a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios, as proposed in this invention. Detailed Implementation

[0035] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0036] Figure 1 This is a structural block diagram of a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios according to the present invention. The device includes:

[0037] Evaluation Updater: Used to evaluate key-value pairs consisting of input and prediction results according to fixed rules or human intervention. There are two evaluation methods: one is manual evaluation, which interacts with the user to ask about the user's true needs; the other is data-driven evaluation, which evaluates entirely based on the results of historical data. If the evaluation result is positive (user satisfied or data correct), the resource manager is updated, and the corresponding input-prediction result is stored in the resource manager in plain text form (the maximum index of the resource storage is incremented by one and stored at that index location). If the evaluation result is negative, (under the manual evaluation method) the user's expected result is asked and the result is stored in the resource storage; (under the data-driven method) the content related to the data is directly deleted from the resource storage.

[0038] Resource Memory: Used to store experiential data (with time indexes, index numbers from 0 to N; even if the evaluation updater deletes a data in the resource memory, the index number will always be a consecutive natural number, and the new data obtained by the evaluation updater will be placed at index N+1). It is also used to guide the intuition generator and the training updater. The dashed arrow indicates that the resource memory can also modify the number of output nodes of the cognitive model in the cognitive device by traversing the data types.

[0039] Training Updater: The data (input-prediction) in the resource store is packaged into pairs (12 data pairs per data frame). After processing by the built-in encoder, the data is used to train and update the network weights of the cognitive model (neural network) in the cognitive device, using the minimum cross-entropy error and gradient descent principle (see pseudocode for details). Figure 4 Training for 300 epochs. The cross-entropy error formula is shown in Formula 1 (where x is the input, y is the output, ω is the weight, and n represents the index number corresponding to the variable). This represents the connection between the nth input node and the nth y, only when... (The entire error matrix is ​​multiplied by the identity matrix only if there are no unnecessary nodes)

[0040] (1)

[0041] Encoder: Used to encode plain text objects, converting plain text into a matrix composed of word vectors (encoded by "-1", "0", "1") that can be computed by neural networks (in most cases, the text object after transcoding is not enough to fill the entire 32*32 matrix, so the problem is solved by copying the transcoded data to fill the matrix by dividing the matrix into equal parts).

[0042] Decoder: Used for decoding the prediction results of neural networks. It decodes the prediction results (a dictionary consisting of category numbers and corresponding results, and the non-linear probability corresponding to each number) into a list consisting of the text format of the object with the highest non-linear probability and its non-linear probability.

[0043] The Cognitor consists of a cognitive model and a prior model (a neural network trained with prior knowledge). It simulates the human cognitive process, expanding or shrinking the output nodes of the cognitive model (a neural network with a Softmax layer can output non-linear probabilities) based on the number of object types in the resource memory. It also identifies the predicted value corresponding to the encoder's output, which is then processed by the decoder to obtain the corresponding output (a list of outputs and non-linear probabilities). The Cognitor also has a special mechanism: when the number of nodes in the cognitive model changes, the weights and structure of the trained cognitive model are overlaid on the prior model. Aside from the variable output nodes, the remaining network structure consists of three convolutional pooling groups and one fully connected layer. The activation function for the first six layers is sigmoid, while the last layer uses Softmax.

[0044] The Intuition Generator, composed of a habit model and a short-term memory model, simulates human intuition. Both outputs are lists of predictions and nonlinear probabilities. The habit model assigns a nonlinear probability P1 to the prediction based on the distribution of the most frequent prediction in the resource memory (habit parameter σ, as shown in Formula 3, where C represents the number of objects in the resource memory, with the subscript "selected" indicating selection, and n represents the number of object types in the resource memory), as shown in Formula 2, where "outdim" represents the number of object types in the resource manager. The short-term memory model assigns a nonlinear probability P2 to the prediction based on the position of the corresponding prediction on the time axis in the resource memory (distribution parameter τ, as shown in Formula 5, where I represents the time index number, with subscripts "max" and "selected" indicating the largest and selected values, respectively), as shown in Formula 4. If there is no corresponding prediction in the resource memory, it means the framework has no memory of this output, i.e., the short-term memory model has no output.

[0045] (2)

[0046] (3)

[0047] (4)

[0048] (5)

[0049] Decision Maker: It aggregates the outputs of all modules and selects the object with the highest non-linear probability as the final output.

[0050] This implementation uses a human-like cognitive reasoning framework for predicting user needs in elderly care scenarios to predict user needs, such as... Figure 2 As shown, it includes the following steps:

[0051] Step 1: Based on the historical data set in the resource storage, train the cognitive model in the cognitive device using the training updater, and load the weights of the prior model in the cognitive device. It is worth noting that the training updater will only train the prior model in the cognitive device on the first use, and the weights of the prior model are obtained from the initial training when loading the weights.

[0052] Step 2: The input is processed by the encoder-cognitor-decoder to obtain the result, and the intuition generator produces the corresponding result.

[0053] Step 3: Input the results obtained in Step 2 into the decision maker to obtain the final prediction results of the framework.

[0054] Step 4: Input the final input and prediction results into the evaluation updater to determine if the result is positive. If it is, proceed to step 5; otherwise, proceed to step 6.

[0055] Step 5: Store the positive evaluation data into the resource storage.

[0056] Step 6: Determine if the non-positive result exists in the resource storage. If it does, proceed to Step 7; otherwise, proceed to Step 8.

[0057] Step 7: Delete the result from the resource storage.

[0058] Step 8: Query the expected result and store the result in the resource storage.

[0059] Step 9: Train the updater to perform operations to train the cognitive model in the cognitive unit.

[0060] Step 10: Determine whether the output node of the cognitive model has changed. If it has, proceed to step 11; otherwise, proceed to step 12.

[0061] Step 11: The cognitive model in the cognitive device performs a coverage operation on the prior model.

[0062] Step 12: Determine if there is new input. If so, repeat steps 1 to 12; otherwise, terminate the process.

[0063] A human-like cognitive reasoning framework for predicting user needs in an elderly care context was used to perform a task of inferring a natural person's need for fruit based on environmental and physical conditions, in order to verify the effectiveness of the proposed method. The results are as follows: Figure 3 As shown. The requirements are divided into three features: two descriptive features and one object, such as "cool-sweet-watermelon". The input is plain text describing the external environment and the person's physical state (sometimes only partial conditions are provided for generalization purposes, and the preferences of a specific natural person are continuously fitted in subsequent manual evaluation).

[0064] The dataset consists of a natural person's demand for fruit under certain conditions, recorded in chronological order (using a questionnaire). Without altering the chronological order, the data is divided into 20 groups of 50. The data is then sequentially entered into a framework, and the data provider (in their physiological and psychological state at the time of the 20th data point) judges its accuracy (Data Source Validation, DSV) and uses this as an evaluation metric for manual assessment. The accuracy calculation formula for group "i" is shown in Formula 7, where "state" represents the DSV status as shown in Formula 6.

[0065] (6)

[0066] (7)

[0067] Depend on Figure 3 The trend of the prediction accuracy results shown demonstrates that the present invention continuously improves its understanding by following human preferences and responds readily to new demands, thus meeting the requirements of the method to enhance human-computer interaction experience, reduce the number of inquiries, and dynamically update to keep up with human changes, proving the effectiveness of the method.

[0068] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. The scope of the present invention is defined only by the appended claims.

Claims

1. A human-like cognitive reasoning system for predicting user needs in elderly care scenarios, characterized in that... include: Resource storage: Used to store reliable data pairs, record established facts and predict successful cases, and also for counting and accessing cases and objects; Encoder: Used to encode plain text objects, transforming plain text into a matrix of word vectors that can be computed by neural networks; Decoder: Used for decoding the prediction results of neural networks, decoding the prediction results into a list consisting of the text format corresponding to the object with the highest non-linear probability and its non-linear probability; Evaluation updater: Used to integrate inputs and system outputs for evaluation. There are two evaluation methods: one is manual evaluation, in which the system interacts with the user to ask for the user's true needs; the other is data-driven evaluation, which is based entirely on the results of historical data. If the evaluation result is positive, the data is stored in the resource storage; otherwise, the resource storage is traversed. If the same data is found, the data is deleted or the data obtained from the user is stored, and the time index of the resource storage is updated. Intuition Generator: Composed of a habit model and a short-term memory model, it is used to simulate human intuition. The output of both is a list of prediction results and nonlinear probabilities. The habit model assigns a nonlinear probability to the prediction result based on the distribution of the most frequent prediction result in the resource memory. The short-term memory model assigns a nonlinear probability to the prediction result based on the distance of the corresponding prediction result in the resource memory on the time axis. If there is no corresponding prediction result for the input in the resource memory, it means that the system has no memory of this output, that is, the short-term memory model has no output. Cognitive device: Composed of a cognitive model and a prior model, it is used to simulate the human cognitive process. It expands or reduces the output nodes of the cognitive model according to the number of object types in the resource storage. It is also used to identify the predicted value corresponding to the data output of the encoder. The predicted value is processed by the decoder to obtain the corresponding output. The cognitive device also has a special mechanism that when the number of nodes of the cognitive model changes, the weights and structure of the trained cognitive model are overlaid on the prior model. Training updater: The built-in encoder is used to package the data in the resource storage, encode it, and then use it to train and update the network weights of the cognitive model in the cognitive device. Decision Maker: It aggregates the outputs of all modules and selects the object with the highest non-linear probability as the final output.

2. The human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... The data pair is "input-prediction result", and the accessed records have time index numbers ranging from 0 to the maximum.

3. The human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... The word vectors are encoded using "-1", "0", and "1".

4. The human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... If the encoded text object is insufficient to fill the entire 32*32 matrix, the encoder will copy the encoded data to fill the matrix by dividing the matrix into equal parts.

5. A human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... The decoder predicts results as a dictionary consisting of category numbers and corresponding results, and a nonlinear probability for each category number.

6. The human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... The evaluation updater outputs a prediction result, with a positive evaluation result indicating user satisfaction or data accuracy.

7. The human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... In the cognitive device, the prior model refers to a neural network trained with prior knowledge, the cognitive model refers to a neural network with a Softmax layer that can output nonlinear probabilities, and the output refers to a list composed of the output and the nonlinear probabilities.

8. A human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... In the training updater, data refers to the input-prediction result, and the cognitive model refers to the neural network.

9. A human-like cognitive reasoning system for predicting user needs in elderly care scenarios according to claim 1, characterized in that... A method for predicting a natural person's demand for fruit using a human-like cognitive reasoning system for predicting user needs in an elderly care context includes the following steps: Step 1): Train the cognitive model in the cognitive device using the training updater based on the historical data set in the resource storage, and load the weights of the prior model in the cognitive device. Step 2): Use an encoder to transcode the input plain text external conditions, and then input the transcoded data into the cognitive device; Step 3): Input the outputs of the cognitive model and the prior model in the cognitive device into the decoder to obtain two lists consisting of the prediction results and nonlinear probabilities representing their occurrence probability; Step 4): Based on the distribution of historical data in the resource storage, find the prediction result with the highest frequency and assign it a non-linear probability P1 as the output of the habit model in the intuition generator. If there is a corresponding prediction result in the input resource storage, assign that result as the output of the short-term memory model in the intuition generator and assign it a non-linear probability P2. Step 5): Integrate the outputs of the cognition device and the intuition generator into the decision-making device to obtain the final prediction result; Step 6): Input the input and the final prediction result into the evaluation updater. If the evaluation result is positive, proceed to step 7); otherwise, proceed to step 8. Step 7): Increment the maximum time index number of the resource manager by one, and store the input and the final prediction result at the maximum time index of the resource storage; Step 8): Search the resource manager. If a data set consisting of the input and the final prediction result exists in the resource manager, delete the data and update the time index number. Step 9): If a new object is input, repeat steps 1) through 8).

Images