53 results about "Strategic learning" patented technology

Systems and methods are provided autonomous driving policy generation. The system can include a set of autonomous driver agents, and a driving policy generation module that includes a set of driving policy learner modules for generating and improving policies based on the collective experiences collected by the driver agents. The driver agents can collect driving experiences to create a knowledge base. The driving policy learner modules can process the collective driving experiences to extract driving policies. The driver agents can be trained via the driving policy learner modules in a parallel and distributed manner to find novel and efficient driving policies and behaviors faster and more efficiently. Parallel and distributed learning can enable accelerated training of multiple autonomous intelligent driver agents.

The invention provides a fictitious self-play-based multi-person incomplete information game policy resolving method, device and system as well as a storage medium. The method comprises the followingsteps: specific to a two-person gaming condition, implementing the generation of an average policy by using multi-type logistic regression and reservoir sampling, and implementing the generation of anoptimal response policy by using a DQN (Deep Q-Network) and annular buffering memory; and specific to a multi-person gaming condition, implementing the optimal response policy by using a multi-agentproximal policy optimization (MAPPO) algorithm, and meanwhile adjusting agent training by using multi-agent NFSP (Neural Fictitious Self-Play). The method has the beneficial effects that a fictitiousself-play algorithm framework is introduced; the Texas Poker policy optimizing process is partitioned into optimal response policy learning and average policy learning which are implemented by simulation learning and deep enhancement learning respectively; and a more universal multi-agent optimal policy learning method is designed.

Efficient exploration of natural language conversations associated with dialogue policy learning may be performed using probabilistic distributions. Exploration may comprise identifying key terms associated with the received natural language input utilizing the structured representation. Identifying key terms may include converting raw text of the received natural language input into a structured representation. Exploration may also comprise mapping at least one of the key terms to an action to be performed by the computer system in response to receiving natural language input associated with the at least one key term. Mapping may then be performed using a probabilistic distribution. The action may then be performed by the computer system. A replay buffer may also be utilized by the computer system to track what has occurred in previous conversations. The replay buffer may then be pre-filled with one or more successful dialogues to jumpstart exploration.

The invention provides an AGV path planning method and system based on reinforcement learning, and aims to solve the problem that an existing AGV path planning method based on reinforcement learning needs to consume a large amount of time and computing power cost. According to the method, an AGV dynamical model is constructed; an AGV is used as an intelligent agent, environment information sensed by the driving of the AGV is used as state information, a state space is designed with a destination position and an obstacle position considered, and a continuous action space and a multiple reward mechanism are designed; markov process modeling of path planning is completed on the basis of the state space, the continuous action space and the multiple reward mechanism, any different starting points, target points and obstacles at any positions can be given in the state space, generalization is high; an Actor-Critic framework is introduced, strategy learning training is carried out, online operation avoids the problem of large calculation amount; and calculation requirements are low, and the real-time decision control of the AGV for any target and obstacle can be realized.

The invention relates to an intelligent agent control method and device based on reinforcement learning. The method comprises the steps of obtaining current local observation of an intelligent agent;taking the current local observation of the intelligent agent as the input of a reinforcement learning model, and obtaining the current execution action of the intelligent agent output by the reinforcement learning model; controlling an intelligent agent to execute the current execution action of the intelligent agent; according to the technical scheme provided by the invention, the strategy learning process in a large-scale multi-intelligent system can be effectively simplified, the number and types of intelligent agents are easy to expand, and the method has potential value in large-scale real world application.

The invention discloses a joint learning method for features and strategies based on state features and subsequent features. The method comprises: obtaining state features representing an input stateby learning mapping from the input state to the instant award; by learning mapping from state features to a value evaluation function, obtaining subsequent features; the obtained state features and the subsequent features being on different time resolution ratios, after the state features and subsequent features are fused, learning fusion results by using a variety of strategic learning networks.Compared with a traditional Agent network, the method uses sample information more efficiently. Compared with other algorithms, learning speed is obviously accelerated, and a network can converge faster and obtain better learning effect.

The invention relates to a network access control method and device. The method comprises the following steps: receiving a strategy obtaining instruction, wherein the strategy obtaining instruction isused for indicating a first terminal to obtain a target network strategy from the trusted management server; wherein the target network strategy is used for controlling first network data on the first terminal, the target network strategy is generated through a first strategy learning process executed on a second terminal, and the first strategy learning process is a process of learning a first access log on the second terminal; obtaining a target network strategy from the trusted management server in response to the strategy obtaining instruction; sending strategy effective information to atrusted management server, wherein the strategy effective information is used for indicating that the target network strategy is confirmed to be effective on the first terminal; and executing the target network strategy to control the first network data on the first terminal. According to the application, the technical problem that the configuration efficiency of the network access strategy is relatively low in the related art is solved.

The invention discloses a multi-vehicle collaborative planning method based on crowd-sourcing learning, and belongs to the technical field of multi-vehicle-road collaborative decision making. According to the invention, the edge server is utilized to reduce the requirements of the computing capability and the communication capability of the vehicle; the evolutionary game is used for modeling the process of continuous game between vehicles in routing planning, and when the game state forms a stable situation, each vehicle obtains a routing decision with maximum own benefit; an intersection passing driving decision-making module is deployed on each vehicle, the vehicle is regarded as an independent decision-making individual, and a cooperative driving behavior of multiple vehicles at the intersection is modeled by using the powerful strategy learning capability of deep reinforcement learning; a traffic situation prediction module is calculated and deployed at the roadside edge, and the traffic situation perception under the limited visual field of vehicles is expanded by using the communication capability of multiple vehicles and roads. According to the invention, different aspects of road resources are optimized, space-time utilization of the intersection is optimized, space-time utilization of road resources around the intersection is optimized, and throughput of the intersection is increased.

The invention discloses a student learning emotion recognition method based on a convolutional neural network. The method comprises the following steps: classifying the expressions of a student through a convolutional neural network model, dividing the learning emotion of the student into positive emotion and negative emotion according to the expressions, storing the student information and emotion information, and feeding back the learning emotion of the student to a teacher. Parents and students can solve the problem that emotions of the students are not easy to perceive, support is provided for teachers to optimize classroom settings and pay attention to learning emotions of the students, a positive role is played in guaranteeing the classroom effect, support can be provided for online learning and detection of student input degree, and an educator can adjust teaching strategies and a learner can adjust learning states.

The invention discloses a spoken language dialogue management system which comprises a voice recognizer, a language understanding unit, a dialogue manager, a language generation unit, a voice synthesizer, a context guidance information generator and a strategy learning unit. In addition, the invention further discloses a spoken language dialogue management method. By the adoption of the technicalscheme, the dialogue manager maintains dialogue state information in the dialogue process, the system generates context guidance information according to the current dialogue state information and dynamically guides the voice recognizer to better recognize natural language possibly used by a user in the current context, and the voice recognition accuracy is greatly improved; and the speech recognizer does not need to prepare training corpora in a specific field in advance and train language models related to the field, but dynamically adjusts the language models according to different contexts, so that the workload of training the language models in advance is reduced, and the production efficiency of a man-machine conversation system is improved.

The invention belongs to the technical field of market transaction, and particularly relates to an optimal bidding strategy solving method. The traditional evolutionary game method is difficult to obtain a stable evolutionary equilibrium solution when facing the uncertainty of opponent decision. The invention provides an optimal bidding strategy solving method. The method aims at providing a DR resource optimal bidding strategy solving problem for users participating in demand side bidding, and comprises the steps of establishing a user participation DR income model based on a consumer demand curve. comprehensively considering the target orientation of the user participating in the market, and proposing an evolutionary game model based on the bounded rationality of a game main body; and 2) considering the incompleteness of market information and the uncertainty of subject decision, and providing an optimal bidding strategy learning algorithm based on Q learning and compound differential evolution for solving the established evolutionary game model based on the finite rationality of the game subject. An evolutionary game model based on user bounded rationality is established, and the user is helped to formulate an optimal DR bidding strategy.

The embodiment of the invention provides a self-learning credible strategy construction method and system based on SELinux. The method comprises the steps: acquiring an application program needing tobe added into a strategy file; entering a strategy learning mode, and setting the SELinux as a tolerance mode; allowing and executing all operations of the application program, and obtaining rejectioninformation recorded in a log file; reading rejection information, and converting the rejection information into a strategy file; and loading the strategy file to the kernel. According to the self-learning credible strategy construction method and system, SELinux is to be in a tolerance mode for the newly installed application program, the strategy file is constructed through one training process, the application program operation needing to be added into the strategy file is executed, under the condition that the front operation strategy is not added, the rear operation is still allowed to be executed, the learning process cannot be stopped, and the credible strategy needed by the application program is constructed under the condition that program running is not affected.

The invention discloses an information physical system security control method based on deep reinforcement learning, and belongs to the technical field of information security. According to the invention, the problem of poor control performance of a security control strategy designed based on an existing method under the condition of network attack is solved. According to the method, the dynamic equation of the cyber-physical system under the attacked condition is described as a Markov decision process, and based on the established Markov process, the security control problem of the cyber-physical system under the false data injection attack condition is converted into a control strategy learning problem only using data; based on a flexible action-critic reinforcement learning algorithm framework, a flexible action-critic reinforcement learning algorithm based on a Lyapunov function is proposed, a novel deep neural network training framework is provided, a Lyapunov stability theory is fused in the design process, the stability of an information physical system is ensured, and the control performance is effectively improved. The method can be applied to safety control of the information physical system.

A controller for controlling a system that includes a policy configured to control the system is provided. The controller includes an interface connected to the system, the interface being configured to acquire an action state and a measurement state via sensors measuring the system, a memory to store computer-executable program modules including a model learning module and a policy learning module, a processor configured to perform steps of the program modules. The steps include offline-modeling to generate offline-learning states based on the action state and measurement state using the model learning program, providing the offline states to the policy learning program to generate policy parameters, and updating the policy of the system to operate the system based on the policy parameters. In the policy learning program to generate the policy parameters are considered a dropout method to improve the optimization of the policy parameters, a particle method to compute and evaluate the evolution of the particle states and a model of the sensor and a model of the online estimator to generate particle state online estimates to approximate the state estimates based on the particles states generated from the model learning program.

The invention relates to a multi-task network model training method and system based on adaptive task weights. According to the method, a sharing mode is learned through a strategy specific to tasks, the strategy autonomously selects which layers are executed in a multi-task network, and weights matched with the tasks can be searched at the same time, so that the model is better trained. According to the method, the multi-task network model is reconstructed based on ResNet, the learning strategy is effectively optimized according to the image in the data set in the training process, and the oneness of the multi-task model is overcome while the task index is improved. According to the method, a multi-task loss function suitable for regression and classification tasks is deduced based on probability theory maximum likelihood estimation, the task weight can be automatically adjusted in the training process so as to better improve the model performance, and the problem that the task weight is not flexible is solved.

The invention discloses a Ceph system performance optimization system based on deep reinforcement learning. The Ceph system performance optimization system is composed of a data source module, a dataaccess mode learning module, an evaluation mechanism learning module and a system parameter adjustment learning module. The Ceph system performance optimization strategy based on deep reinforcement learning is realized through the following steps: S1, preprocessing a data source; s2, learning and classifying a Ceph file system running environment model; s3, carrying out evaluation mechanism learning; and S4, learning a Ceph file system parameter adjustment strategy. According to the data access method, deep reinforcement learning algorithm and interactive learning of an A2C model and a Ceph file system are combined to obtain the optimized parameters, and optimal system parameters adapted to data access mode may be selected; the method can adapt to different data access modes and hardware configurations, the optimal system parameters are obtained through intelligent learning, the system parameters can be obtained according to the optimal system parameters, and therefore the performanceof the Ceph file system is improved.

Sample utilization rate and learning efficiency are an important bottleneck problem of deep reinforcement learning in practical application. For the real world, in order to quickly and accurately obtain a universal strategy, the invention provides a reinforcement learning strategy learning method based on a potential action representation space, and the method learns the strategy in the potentialspace of the action and maps the action representation to the real action space; a strategy in the method is mapping from a state to action representation, the search space of strategy learning can bereduced, and the strategy learning efficiency is improved; according to the method, mature supervised learning can be selected for offline learning for representation of actions, the learning speed can be further increased, and the stability is improved. Besides, as long as the characteristics of the adopted actions are similar, even for tasks different from the training strategy, the learned strategy can be generalized into the action space of the current execution task under the fine tuning of a small number of learning samples, and the generalization ability of strategy expression is greatly improved.

The invention relates to a task-oriented dialogue strategy generation method, and the method comprises the following steps: establishing a dialogue state tracker, and determining a dialogue state space, an action space and formalized representation thereof; simulating a dialogue state by using a variational automatic encoder; simulating a dialogue action by using a multi-layer perceptron and Gumbel Softmax; performing adversarial training on a simulation sample generator and a discriminator; and finally training a dialogue strategy by using a reinforcement learning method. Firstly, a simulation sample generator is used for learning a reward function, and loss from a discriminator can be directly fed back to the generator for optimization; secondly, the trained discriminator is taken as a dialogue reward to be brought into a reinforcement learning process for guiding dialogue strategy learning; the dialogue strategy can be updated by utilizing any reinforcement learning algorithm; according to the method, common information contained in high-quality dialogues generated by human beings can be deduced by distinguishing the dialogues generated by the human beings and the machine respectively, and then the learned information is fully utilized to guide dialogue strategy learning in a new field in a transfer learning mode.

The invention discloses a large-scale unmanned aerial vehicle cluster flight method based on deep reinforcement learning, and the method comprises the steps: dividing a learning process of an unmanned aerial vehicle cluster anti-collision strategy into a plurality of courses in sequence, and enabling the unmanned aerial vehicle cluster scale of the next course to be larger than the unmanned aerial vehicle cluster scale of the previous course; constructing a curriculum reinforcement learning framework based on an actuator network and an evaluator network, and setting a group constant network based on an attention mechanism in the curriculum reinforcement learning framework; sequentially carrying out strategy learning on each course according to the course reinforcement learning framework to obtain a flight strategy of each unmanned aerial vehicle; and according to the empirical data of each unmanned aerial vehicle in the previous course of the current course, the executor network parameters and the evaluator network parameters of the current course in the strategy learning process are updated. According to the invention, the learning and training efficiency of the large-scale unmanned aerial vehicle can be effectively improved, the collision of the large-scale unmanned aerial vehicle cluster during flight is effectively avoided, and the generalization ability is strong.