80 results about "Reinforcement learning control" patented technology

The invention, which belongs to the technical field of artificial intelligence and control, relates to a neural network and enhanced learning algorithm, particularly to an inverted pendulum control method based on a neural network and reinforced learning, thereby carrying out self studying to complete control on an inverted pendulum. The method is characterized in that: step one, obtaining inverted pendulum system model information; step two, obtaining state information of an inverted pendulum and initializing a neural network; step three, carrying out and completing ELM training by using a straining sample SAM; step four, controlling the inverted pendulum by using an enhanced learning controller; step five, updating the training sample and a BP neural network; and step six, checking whether a control result meets a learning termination condition; if not, returning to the step two to carry out circulation continuously; and if so, finishing the algorithm. According to the invention, a problem of easy occurrence of a curse of dimensionality in continuous state space as well as a control problem of a non-linear system having a continuous state can be solved effectively; and the updating speed becomes fast.

The invention discloses a space robot mechanical arm arresting control system. The space robot mechanical arm arresting control system comprises two loops, namely, an inner loop and an outer loop, inthe outer loop, the system achieves the attitude stability of a space robot mechanical arm base platform in the arresting process through a PD controller, and in the inner loop, the system controls amechanical arm to achieve arresting maneuvering on a non-cooperative target through a reinforce learning control system based on reinforce learning. The invention further discloses a reinforce learning method for controlling the reinforce learning control system of the mechanical arm in the inner loop of the system and a space robot dynamics modeling method of the space robot mechanical arm arresting control system. According to the space robot arresting control system, the reinforce learning method and the dynamics modeling method, compared with PD control, the posture disturbance of the baseplatform under reinforce learning RL control is smaller, the movement process of the tail end of the mechanical arm is more stable, the control precision is higher, moreover, the motion flexibility of the mechanical arm under the reinforce learning RL control is good, and the autonomous intelligence is achieved to the greater extent.

The invention relates to a quad-rotor unmanned aerial vehicle reinforcement learning nonlinear attitude control method, and aims to solve a problem of quad-rotor unmanned aerial vehicle attitude control of a quad-rotor unmanned aerial vehicle dynamical model with an unmodeled part; and a reinforcement learning controller based on an execution-evaluation neural network is designed for estimating the unmodeled part of the model. And meanwhile, a nonlinear robust controller based on multivariable super-twisting is designed, so that stable attitude control of the quadrotor unmanned aerial vehicleis realized.

The invention provides a deep reinforcement learning control method for vertical path following ofan intelligent underwater robot. The deep reinforcement learning control method comprises the following steps that firstly, according to path following control requirements of the intelligent underwater robot, an intelligent underwater robot environment is established to interact with an agent; secondly, an agent set is established; thirdly, an experience buffer pool is established; fourthly, a learner is established; and fifthly, intelligent underwater robot path following control is conducted byusing a distributed deterministic strategy gradient. According to the deep reinforcement learning control method for the vertical path following ofthe intelligent underwater robot, the deep reinforcement learning control method for the vertical path following ofthe intelligent underwater robot is designed to solve the problem that marine environment in which the intelligent underwater robot is located is complex and variable, thus a traditional control method can not interact with the environment. The path followingand control task of the intelligent underwater robot can be finished in a distribution mode by using a deterministic strategy gradient, and the deep reinforcement learning control method for the vertical path following ofthe intelligent underwater robot has the advantages of self-learning, high precision, good adaptability and stable learning process.

The invention relates to a heavy haul train reinforcement learning control method and system, and relates to the technical field of heavy haul train intelligent control, and the method comprises the steps: obtaining the running state information of a heavy haul train at a current moment; according to the running state information of the heavy haul train at the current moment and a heavy haul trainvirtual controller, obtaining a heavy-haul train control instruction at a next moment, and sending the heavy haul train control instruction at the next moment to a heavy haul train control unit to control the heavy haul train to run; the heavy haul train virtual controller is obtained by training a reinforcement learning network according to heavy haul train running state data and an expert experience network, wherein the reinforcement learning network comprises a control network and two evaluation networks, and the reinforcement learning network is constructed according to an SAC reinforcement learning algorithm. According to the invention, the heavy haul train can have the properties of safety, stability and high efficiency in the running process.

The invention relates to a blind guiding robot of an optimal output feedback controller and based on reinforcement learning, and belongs to the technical field of robots. By adopting a realsense D435idepth camera as a visual sensor, real-time environment information of the blind guiding robot in the advancing guiding process can be accurately and efficiently obtained. In order to solve the problem that a blind guiding robot faces many unstable factors in the moving process, a model-free synchronous integral reinforcement learning controller based on an ADP method is designed, an HJB equationof a constructed cost function is established by constructing the cost function of a blind guiding robot system based on reinforcement learning, the HJB equation is solved through a method based on synchronous reinforcement learning, and finally an optimal solution is obtained through an iterative method, so that the optimal control of the blind guiding robot system is achieved.

The invention discloses a microgrid hybrid coordination control method based on reinforcement learning and a multi-agent theory. The method comprises the steps of designing a transition voltage layercontrol strategy based on a voltage layering mode, designing a double-energy-storage sub-role control strategy, and enabling two energy storage to work separately when an energy storage unit works ina voltage stabilization mode; when energy storage assistance is needed to continuously absorb power or supplement power, converting the two-energy-storage working mode into cooperative charging/discharging; constructing an action space and a state space based on Q-Learning; designing a reinforcement learning control framework based on multiple agents, specifically, designing a basic updating ruleof a state-action pair and selecting a corresponding value function; designing a basic action selection mechanism and a return value strategy, specifically, designing a selection strategy adopted by the system in an initial state and a return value in each state; and designing a reinforcement learning algorithm flow, specifically, designing a proper algorithm flow based on the strategy to realizea control strategy.

The invention discloses a brain-computer cooperation digital twinning reinforcement learning control method and system. A brain-computer cooperation control model is constructed, an operator gives a virtual robot direction instruction, meanwhile, an electroencephalogram signal generated when the operator gives the virtual robot direction instruction is collected, a corresponding speed instructionof a virtual robot is given according to the collected electroencephalogram signal to complete a specified action, reward value calculation is performed on the brain-computer cooperation control modelaccording to the completion quality, training of the brain-computer cooperation control model is completed, a double-loop information interaction mechanism between the brain and the computer is realized through a brain-computer cooperation digital twinning environment in a reinforced learning manner, and interaction of an information layer and an instruction layer between the brain and the computer is realized. According to the method and the system, the brain state of the operator is detected through the electroencephalogram signals, compensation control is conducted on the instruction of the robot according to the brain state of the operator, accurate control is achieved, and compared with other brain-computer cooperation methods, the method has the advantages that the robustness and generalization ability are improved, and mutual adaptation and mutual growth between the brain and the computers are achieved.

The invention provides a vector reinforcement learning control method of a power grid frequency modulation type flywheel energy storage system. The method can overcome the current situation that traditional frequency modulation resources cannot meet frequency modulation requirements due to impact of randomness, volatility and uncertainty of new energy power generation and distributed power generation on a power grid in an existing power system. Frequency modulation of the flywheel energy storage system is combined with vector reinforcement learning, the optimal action of the flywheel energy storage system is selected by performing vector reinforcement learning on voltage, a motor of the system is controlled to work in a generator/motor state to achieve the purpose that the system works ina discharging/charging mode, and therefore the purpose of adjusting the frequency of a power system is achieved. According to the vector reinforcement learning control method of the power grid frequency modulation type flywheel energy storage system, the response speed is much higher than that of a traditional frequency modulation resource, the power grid frequency can be rapidly adjusted, the power grid frequency is kept within the allowable deviation range, the system frequency stability is maintained, and therefore the reliability and safety of power grid operation are guaranteed.

The invention discloses an automobile longitudinal multi-state control method based on deep reinforcement learning priority extraction. The method comprises the following steps: 1, defining a state parameter set s and a control parameter set a of automobile driving; 2, initializing deep reinforcement learning parameters, and constructing a deep neural network; 3, defining a deep reinforcement learning reward function and a priority extraction rule; 4, training the deep neural network and obtaining an optimal network model; and 5, obtaining a t moment state parameter st of the automobile, inputting the t moment state parameter st into the optimal network model to obtain an output at, and executing the output at by the automobile. According to the method, the multi-state driving of the automobile in the longitudinal direction is completed by combining the priority extraction algorithm and the deep reinforcement learning control method, so that the safety of the automobile in the driving process is higher, and traffic accidents are reduced.

The invention provides a quantum deep reinforcement learning control method of a doubly-fed wind generator. The method can solve the control problem of stator flux linkage change of the doubly-fed wind generator after a power grid fault is removed and when the power grid voltage is asymmetrically and suddenly increased. The method is a control method combining Bayesian quantum feedback control, deep learning and reinforcement learning. The Bayesian quantum feedback control process is divided into two steps of state estimation and feedback control, and feedback input is historical measurement and current measurement records. Bayesian quantum feedback can effectively control decoherence in solid quantum bits. The deep learning part adopts a convolutional neural network model and a back propagation method. In the reinforcement learning part, Q learning based on a Markov decision process is used as a control framework of the whole method. According to the method, the control stability of the doubly-fed wind generator can be effectively improved, and the wind energy utilization efficiency is improved.

The invention discloses a quasi-proportional resonance controller parameter adjustment method. The method comprises the following steps of: obtaining models of an inverter and a load, and adjusting areinforcement learning training environment by taking the models as parameters; constructing a depth deterministic strategy gradient reinforcement learning framework, and defining depth deterministicstrategy gradient reinforcement learning framework parameters, wherein the deep deterministic strategy gradient reinforcement learning framework parameters comprise a state, an action and a reward value; and training the intelligent agent of the parameter adjustment reinforcement learning framework in the parameter adjustment reinforcement learning training environment. According to the method forsetting control parameters of the multi-parallel quasi-proportional resonance controller based on the reinforcement learning method, because the reinforcement learning control algorithm is not sensitive to the mathematical model and the operation state of the controlled object, the self-learning capability of the reinforcement learning control algorithm has strong adaptability and robustness to parameter change or external interference, the control requirement can be met when the multi-quasi-proportional resonance controller is connected in parallel, and the control effect can be ensured whenthe load is changed.

The invention provides a motion control decision generation method and device, electronic equipment and a storage medium, and the method comprises the steps: determining a depth prediction map corresponding to the RGB image information based on the first point cloud information, obtained by a laser radar, of a target region and the RGB image information, obtained by a camera, of the target region; determining target point cloud information based on second point cloud information of the depth prediction map projected into the three-dimensional space, and obtaining third point cloud information according to the target point cloud information and the first point cloud information; and determining target multivariate state data according to the third point cloud information, and inputting the target multivariate state data into the deep reinforcement learning control decision model after multi-stage training to obtain a target motion control decision. According to the method provided by the invention, the plane limitation existing in radar detection can be effectively overcome, the target motion control decision is efficiently obtained, and a better moving obstacle avoidance effect is realized.

The invention discloses a train cooperative operation control method based on reference deep reinforcement learning, and the method specifically comprises the steps: building a train cooperative operation simulation environment, setting a train safety distance and the like, and calculating the estimated shortest real-time distance of two trains; setting a reward function, and establishing an input dimensionality reduction reinforcement learning algorithm controller; a reference controller is added, when the train meets a reference control strategy condition, a reference control signal is used for replacing a reinforcement learning control signal, and the part of data is used for optimizing a reinforcement learning control strategy; training the network until the global reward of the network is optimal and the control result is combined with the expectation, and considering that the preliminary training of the network is completed; and loading the reference control strategy and the reinforcement learning control strategy on the real train, and outputting a train control signal according to the real train information to complete the cooperative operation control of the train. According to the method, the optimal strategy training speed is increased, and the robustness of the control strategy in the actual operation process is ensured.