34 results about "Learning automata" patented technology

The invention relates to a development automatic machine with a brain cognition mechanism and a learning method of the development automatic machine, and belongs to the technical field of intelligent robots. The development automatic machine comprises an inner state set, a system output set, an inner operation behavior set, a state transfer equation, a reward signal, a system evaluating function, a system action selecting probability and a Dopamine response differential signal. By means of the development automatic machine and the learning method, a mathematic model high in generalization ability and wide in application range is provided for the system autonomous development process with the learning automatic machine as the basic frame; by means of the method, a sensorimotor system and an intrinsic motivation mechanism are combined, the self-learning and self-adaption capacity of the system is improved, and the intelligence in the true sense is achieved.

The invention provides a hospitalizing resource scoring and recommending method based on a latent factor model. The method comprises the steps that 1, recommended resource item data are acquired, and data filtering and cleaning are conducted; 2, each parameter in a learning automaton training SVD++ model with continuous actions is used; 3, for a user and all resource items in a set, a scoring predicted value of the user to each resource item is calculated based on the latent factor model; 4, a TopN recommendation list is obtained by using an improved sorting algorithm. According to the hospitalizing resource scoring and recommending method based on the latent factor model, a prediction scoring result and correlation degree of a most relevant implicit type are both taken as basis for generating the recommendation list; meanwhile, benchmark coefficients of the resource items are used for reference, it can be guaranteed that the recommended resource items are more likely to attract interest of a patient, and meanwhile it can be guaranteed that the resource items have high enough scores to obtain the love of the patient.

The invention discloses a stochastic learning automata and fuzzy algorithm based high throughput relay selection method. The problem that network performance severely decreases due to the large link losses between sub-nodes and a coordinator in a wireless sensor network can be solved. The method includes the following steps: through the combination of the wireless sensor network and a stochastic learning automata, source nodes can find the best relay to make a system reach to a balance and stability state through a learning manner; relay nodes can perform AF forwarding on received data, and different sensor data has different priority; nodes having the highest priority can access to channels for many times in a frame so that the successful probability of sending can be higher; and the relay nodes adopts a fuzzy algorithm to realize load balancing. According to the embodiments of the method, that the nodes do not need human intervention during operation can be guaranteed, the steady state can be adaptively achieved, the maximization of overall network throughput can be realized, and the method has wide application values.

Provided is a method of optimizing a deep neural network based on learning automata (LA). In the training phase of a deep neural network, by starting from a fully connected initial network structure, weak connections in the network are continuously found and removed in the process of iterative parameter update through gradient descent. Thus, a more sparsely connected network structure with smaller generalization error is obtained, and image classification can be carried out on test samples more accurately. The weak connections are judged by LA through continuous interaction with the neural network in the process of training. By using the idea of reinforcement learning for reference, introducing a learning automaton algorithm to improve the traditional back propagation algorithm and removing redundant connections to reduce network parameters, the classification accuracy of test samples is increased, and the method has stronger ability to prevent over-fitting.

The invention discloses a method for forming a multi-agent distributed union. According to the method disclosed by the invention, when agents accept different tasks, the appropriateness of the capabilities of the agents and the tasks is different; a small union having two agents is used as an ideal basis unit; and, in combination with task earning characteristics, dynamic distribution of weight values is carried out through a Learning Automata algorithm. The condition that expression is carried out in two aspects including the task angle and the agent angle is sufficiently considered; the method accords with the current situation well; the method is more rational; furthermore, concurrent choice of the tasks is realized; and the task distribution speed and effective utilization of agent resources are enhanced.

The invention discloses a heterogeneous directed sensor network node scheduling method oriented to different priority targets. The method comprises the following steps: converting a heterogeneous directed sensor network node scheduling problem into a set coverage problem, and solving the same by adopting a genetic algorithm, realizing the self-adaptive selection of variation parameters and crossover factors in the genetic algorithm by importing a learning automation into the genetic algorithm, thereby improving the convergence capacity and the optimizing capacity of the genetic capacity, and realizing the target of prolonging the life cycle of a network in the premise of guaranteeing the requirement target of the monitoring target. The method has the advantages that the algorithm is good in convergence, string in optimizing capacity and long in network life.

The invention discloses a cognitive and learning method of a cognitive moving system with an inner engine. The cognitive model system to study the basis of automatic machines, including perceived state set, the set of actions, orientation map collection, curiosity, orientation function, learning orientation matrix, the state transition function and knowledge entropy, more than 10 parts. Perception of the current state of the system model; select action mechanism based on the engine; perform an action, state transition occurs; calculated value orientation function; Updates 'Perception - Movement'mapping; the process is repeated until the entropy minimal knowledge or learning time is greater than terminated time. The invention introduces the engine mechanism with active learning environment, not only the system has a strong self-learning and self-organizing ability, and can effectively avoid damaging small probability events, improve the stability of the system, for the establishment of having robot cognitive development capability to provide a strong foundation.

The invention provides a robot attitude path target track optimization cognitive development algorithm CBCLA. The method is divided into eight parts, that is, a limited internal state set, a system output set, an internal operation behavior set, a state transfer equation, an internal state of the system at the moment t, an evaluation function, striatum matrix action selection probability output and dopamine energy respectively, wherein the eight parts can be expressed by an eight-member array: CBCLA={SC, MC, CbA, f, r(t), BGstrio, BGmatrix, SNDPA}. The method simulates neural activity of a sensory motor system of an organism, takes learning automaton as a frame and combines the feature that an intrinsic motivation mechanism drives autonomic learning of the organisms; the cognitive development algorithm is applied to mobile robot path planning and study; and the robot can gradually master motion balance control skills through autonomic learning development under an unknown environment,and realizes real-time tracking of a target.

The invention discloses a network payment fraud detection method based on a self-learning sliding time window. In order to find a more effective implementation scheme for network payment fraud detection, the method comprises the following steps of obtaining a new transaction record detected by a certain user in real time, and extracting features independent of a sliding time window and features dependent on the sliding time window based on the transaction record and a preset sliding time window; inputting the features independent of the sliding time window and the features independent of the sliding time window into the trained random forest classifier model to obtain and return the probability of fraud possibility of the transaction record. According to the present invention, the slidingtime window size is dynamically learned and adjusted through the learning automaton in reinforcement learning, and the defect that a traditional fraud detection system has lagging is overcome.

The invention discloses a random shortest path realization method based on a hierarchical structure learning automaton. The method comprises the following steps: learning automaton deployment: deploying the learning automatons beginning from a source node and ending at a node of a target stage through a dynamical network; initializing: initializing own probability vector of each learning automaton; path selection: selecting nodes layer by lays from a father node so as to form the current path; environment feedback: comparing a cost function of the current path with a mean of the current sampling path to obtain punishment or award; learning process: updating own probability vector according to a learning algorithm by each learning automaton on the selected path; judging the updating ending process layer by layer, and ending the step if the updating is ended, otherwise, updating a father node return path to select the continuous execution.

The invention provides a method and a system for predicting a pocket exchange network link by adopting a learning automaton. The method comprises the following steps: acquiring historical behavior information between node pairs in an opportunity network; performing type division according to the historical behavior information and the connection frequency degree of the node pairs so as to divide the node pairs in the opportunity network into active node pairs or inactive node pairs; constructing a network link prediction model based on the learning automaton according to the active node pair and the inactive node pair; and operating the network link prediction model, and outputting a prediction result to predict whether a connection is generated between the node pairs in the opportunity network. According to the invention, the network link prediction model is correspondingly constructed based on the historical behavior information between the node pairs in the opportunity network; therefore, the possibility of connection between the node pairs can be predicted more accurately, limitation caused by prediction based on network topology attributes or related attributes of the nodes is prevented, and support can be provided for an upper routing protocol.

The invention provides a convolutional neural network compression method and system based on a learning automaton and a medium. The method comprises a parameter initialization step of initializing theparameters of the learning automaton; a state value selection step of according to the obtained initialized learning automaton parameters, enabling each learning automaton to select a state value ofthe learning automaton according to a preset behavior selection probability to obtain a state value of each learning automaton; and a network structure updating step of updating the network structureaccording to the obtained state value of each learning automaton, and obtaining the updated network structure. According to the method, the learning automaton idea is innovatively used for screening the optimal convolution kernel set in the convolutional neural network, so that the convolutional neural network can complete the network compression task to the maximum extent under the condition of losing a little of classification precision.

The invention discloses a high-reliability adaptive MAC layer scheduling method. The problem that a large amount of energy consumption is caused by cluster head nodes in a wireless sensor network dueto idle listening is mainly solved. The method comprises the steps of building a model for the wireless sensor network; generating a specific frame format, and embedding a queue occupancy rate and a time delay in a frame control field; initializing an action set, a selection probability set and a feedback set; enabling a coordinator to interact with a surrounding environment by using a learning automaton method, and updating an action and a state of the coordinator; dividing the whole learning process into three stages, namely, an initial stage, an exploration stage and a greedy stage, and adopting corresponding search strategies; evaluating an effect of action and environment interaction, and updating the feedback and selection probability sets; and selecting relevant parameters for determining a duty ratio based on the feedback set, so that adaptive MAC layer scheduling is achieved. The adaptive adjustment of the duty ratio by the nodes in the operation period is ensured, and the power consumption is minimized; and the method has a wide application value.

The invention relates to the technical field of electric power system planning, in particular to a power grid investment optimization method based on simulation investment benefit analysis and learning automatons. The method comprises the following steps that first, a grey prediction model is used for conducting data prediction on power grid development data; second, related indexes in an investment benefit index system are calculated according to prediction schemes of various obtained investment data through data prediction for the coming year, and then the simulation investment benefit analysis is carried out to obtain simulation investment comprehensive score results; third, the score results obtained through different schemes are optimized based on a multi-scheme optimization strategy of a learning automaton method, and finally the optimal investment scheme is obtained. The power grid investment optimization method based on the simulation investment benefit analysis and the learning automatons is beneficial to optimizing power grid investment and good in using effect.

A global optimization system and method based on continuous action learning automata are provided, wherein the system includes: an initialization module, a behavior selection module, an environment feedback module, an updating module and an output module, wherein: the initialization module initializes parameters of CALA algorithm and inputs results into the behavior selection module to conduct behavior selection, wherein the behavior is fed back through the application of the path environment and enters environment feedback module, thereby obtaining the corresponding environment feedback and the local optimal solution. The updating module updates the algorithm parameters according to the environment feedback, inputs the updated parameters into the behavior selection module to complete an iteration, and improves the smoothing function. The improved smoothing function is introduced into next iteration environment feedback module for performing iteration many times, thereby obtaining theextreme value point finally. The current environment feedback is input into the output module to output the optimal path as the global minimum. The invention is reasonable in design, introduces a smoothing function and adds a slope component for improvement, so that CALA can easily jump out of a local minimum solution, and the subsequent searching has directionality, thereby greatly improving theconvergence speed and the correct rate of the algorithm.

The invention provides a practical reinforcement learning automaton method for quotation optimization of a power generator under limited information. The method comprises the following steps: S1, initializing an action space probability density function and a historical income cache region of the power generation capacity of the power generator; S2, discretizing the probability density function into a discrete probability density function to obtain a plurality of sub-intervals, selecting an action corresponding to the sub-interval where the random number is located according to the cumulativeprobability of the sub-intervals, and submitting the selected action; S3, evaluating the environmental feedback, calculating clearing income, executing enhanced signal evaluation according to the clearing income, and storing the clearing income into the historical income cache region; S4, updating the discrete probability density function as linear operation of discrete values of the discrete probability density function and the discrete Gaussian neighborhood function at the end points of the subintervals; and S5, judging whether an iteration stopping standard is reached or not, if not, returning to the step S2, and if so, ending the optimization process.

The invention discloses a general distributed graph processing method and system based on reinforcement learning. The method includes: defining a distributed data processing center based on a graph theory to form a distributed graph; utilizing a preset graph cutting model and a preset graph processing model to cut the distributed graph by utilizing a reinforcement learning mode based on a preset constraint condition; allocating a learning automaton to each vertex; finding the most suitable data processing center for the vertex through training, wherein the possibility of each vertex in all data processing centers obeys certain probability distribution, the whole system comprises five steps of action selection, vertex migration, score calculation, enhanced signal calculation and probabilityupdating in each iteration process, and the iteration is judged to be ended when the maximum iteration frequency is reached or the constraint condition is converged. The distributed graph processingmodel formed by the universal distributed graph processing method provided by the invention is a universal distributed graph model, and only different score calculation schemes and different weight vectors need to be designed for different optimization targets.

A learning automaton can be trained to merge data from input data streams, optionally with different data rates, into a single output data stream. The learning automaton can learn over time from the input data streams. The input data streams can be low-pass filtered to suppress data having frequencies greater than a time-varying cutoff frequency. Initially, the cutoff frequency can be relatively low, so that the effective data rates of the input data streams are all equal. This can ensure that initially, high data-rate data does not overwhelm low data-rate data. As the learning automaton learns, an entropy of the learning automaton changes more slowly, and the cutoff frequency is increased over time. When the entropy of the learning automaton has stabilized, the training is completed, and the cutoff frequency can be large enough to pass all the input data streams, unfiltered, to the learning automaton.

The invention relates to a multi-modal optimization system based on a random point positioning algorithm of a learning automaton, and the system comprises an initialization module, a parameter selection module, an environment feedback module, a multi-modal random point positioning optimization module, and an output module. The initialization moduleinitializes system parameters. The parameter selection module performs iterative selection of parameters on each parameter sub-interval in the parameter search space, the parameters are optimized to obtain feedback, the feedback is input into the environment feedback module to obtain environment feedback, and the environment feedback is input into the multi-modal random point positioning optimization module to obtain estimated values of all current optimal parameters; and when the number of iterations in the multi-modal random point positioning optimization module reaches a preset maximum number of iterations, the multi-modal random point positioning optimization module inputs all the obtained optimal parameters to the output module, and the output module outputs an optimal parameter set corresponding to all the optimal parameters. Compared with the prior art, the system has the advantages that all global optimal parameters are found at the same time, and the application range of the random point positioning method is widened.

The invention discloses a directional antenna neighbor discovery method based on a deterministic estimator learning automaton in a D2D network, and the method enables a neighbor discovery process to be modeled as the deterministic estimator learning automaton, and an estimator carries out the approximate estimation of the reward probability of an action according to the previous feedback condition, and strengthens the action with a higher reward probability estimation value. The directional antenna neighbor discovery algorithm based on the Pursuit algorithm is adopted, the direction of the directional antenna is adjusted by updating the sector probability distribution, and the neighbor discovery efficiency is further improved.