118 results about "Action prediction" patented technology

The invention discloses a feature filtering defense method for a deep reinforcement learning model. The feature filtering defense method comprises the following steps that: (1) aiming at the]= DDPG model for generating continuous behaviors, a system comprises an actor network and a critic network, wherein the actor network comprises an action estimation network and an action implementation network, the critic network comprises a state estimation network and a state implementation network, pre-training is conducted on the deep reinforcement learning model DDPG, and a current state, a behavior,a reward value and a next state of a pre-training stage are stored in a cache region; (2) an auto-encoder is trained, feature filtering is performed on an input state by using an encoder of the trained auto-encoder to obtain a feature map corresponding to the filtered input state, and a feature map is stored in the cache region; and (3) a convolution kernel in the pre-trained DDPG model is pruned,action prediction is performed by using the pruned DPG model, and a prediction action is output and executed.

The invention relates to a virtual object control and model training method, a device, a storage medium and equipment, wherein the virtual object control method comprises the following steps: acquiring interactive scene data corresponding to a virtual object to be controlled; carrying out characteristic extraction on the interactive scene data to obtain characteristic data; inputting the characteristic data into a virtual object control prediction model to obtain a target action output by an action prediction layer of the virtual object control prediction model; selecting action parameters output by an action parameter prediction layer corresponding to the target action in a plurality of action parameter prediction layers included in the virtual object control prediction model; controllingthe virtual object to be controlled to execute the target action according to the selected action parameters. The virtual object control and model training method, the device, the storage medium andequipment provided by the invention can provide the control efficiency of the virtual object.

The invention discloses a DBN algorithm based drug targeting protein action prediction method. The method comprises the steps of extracting extended communication fingerprint of the drug by starting with the molecular structure of the drug; extracting tripeptide structural characteristic of the protein by starting with amino acid sequence of the protein, splicing the extended communication fingerprint of the drug and the tripeptide structural characteristic of the protein to compose a drug-protein characteristic vector, then inputting the drug-protein characteristic vector to a depth confidence network, wherein the output of the network is the probability of the drug-protein input by the network for the mutual action, and finally selecting a suitable threshold to judge the pair of the association is set up. The DBN algorithm based drug targeting protein action prediction method can give out possible interaction pair of drug-targeted protein rapidly without manual intervention, can accordingly save drug development and testing costs and speeds up drugging and discovery of new functions of drugs.

The invention provides a display method, a device, equipment and a computer readable storage medium, and the method comprises the steps: obtaining human eye information, and determining a gaze regionand a non-gaze region in display equipment according to the human eye information; determining to-be-displayed content of the gaze area according to the movement parameters of the dynamic object in the gaze area, generating an image of the gaze area, and rendering the non-gaze area to obtain an image of the non-gaze area; and combining the image of the gaze area and the image of the non-gaze areato obtain a combined image, and displaying the combined image on a display device. According to the display method provided by the invention, the fixation area and the non-fixation area in the displayequipment are divided according to the human eye information; by determining the dynamic object in the gaze area and predicting the action of the dynamic object, the content to be displayed in the gaze area can be determined according to the action prediction of the dynamic object, ghosting of the dynamic object during scene movement is eliminated or avoided, the display definition of a dynamic picture is ensured, and the display performance of equipment is improved.

The invention discloses a spatial positioning and clustering based action prediction method and system. The method comprises: performing spatial positioning on target points to obtain position information and motion state information of each frame of the target points; performing information coding according to the position information and the motion state information of at least two frames of continuous target points to establish models of at least two frames of information; clustering the models of the at least two frames of continuous information to obtain a clustering result; establishing a training data set according to the clustering result, and obtaining a working mode according to the training data set; and predicting an action according to the working mode. By implementing embodiments of the invention, the target points are subjected to the spatial positioning and the action is effectively predicted through a clustering technology based on externally input huge data, so that the time delays generated in a signal transmission process are reduced, the subjective delays are effectively reduced, and the user experience is improved.

The invention provides a method for recognizing actions on the basis of deep feature extraction asynchronous fusion networks. The method is implemented by the aid of main contents including coarse-grained-to-fine-grained networks, asynchronous fusion networks and the deep feature extraction asynchronous fusion networks. The method includes procedures of inputting each short-term light stream stackof each space frame and each movement stream of input video appearance stream into the coarse-grained-to-fine-grained networks; integrating depth features of a plurality of action class grain sizes;creating accurate feature representation; inputting extracted features into the asynchronous fusion networks with different integrated time point information stream features; acquiring each action class prediction results; combining the different action prediction results with one another by the deep feature extraction asynchronous fusion networks; determining ultimate action class labels of inputvideo. The method has the advantages that deep-layer features can be extracted from the multiple action class grain sizes and can be integrated, accurate action representation can be obtained, complementary information in a plurality of pieces of information stream can be effectively utilized by means of asynchronous fusion, and the action recognition accuracy can be improved.

The invention discloses a training method of an action control model, and the method is applied to the field of artificial intelligence, and specifically comprises the steps: obtaining first state data corresponding to a target role from a to-be-trained segment; based on the first state data, obtaining an action prediction value through a to-be-trained action control model; determining action dataof the target role according to the action prediction value and the M groups of offset parameter sets; and updating model parameters of a to-be-trained action control model according to the first state data and the action data. The invention further discloses a model training device and a storage medium. According to the method, the predicted value of the joint can be converted into the reasonable motion range of the joint, manual adjustment is not needed, the model training efficiency can be improved, the model training effect can also be improved, and therefore the animation effect of character performance is better.

The present invention is provided with: a player action prediction unit that predicts the action of a player and calculates a sales order information prediction value; a market physical restraint prediction unit that predicts the state of a power transfer path, which includes grid-connection line available capacity indicating excess transferred power between areas, and calculates a prediction value for the state of the power transport path; and a market prediction unit that predicts a market price on the basis of the sales order information prediction value and the prediction value for the power transfer path state.

The invention discloses a data broadcasting system based on user action prediction and a data acceleration method thereof. The method comprises the following steps: A, a set top box sequentially receives page data from pages of all stages and page data corresponding to focuses of all pages in advance and hierarchically stores the page data from pages of all stages and the page data corresponding to focuses of all pages into a buffer; when page browsing is required and user's operational order is received through clicking to a focus, page date corresponding to the focus is called out from the buffer to display. According to the invention, page data possibly browsed through the current action of a user can be received in advance through a multistage page data in-advance receiving mechanism, when a user click a focus, a corresponding page is displayed immediately, a browsing speed is increased, and conveniences are brought to users.

The invention provides a human skeleton-oriented motion prediction method and system. The method comprises the following steps: a data acquisition step: acquiring human skeleton data; a human body multi-scale map construction step: constructing a multi-scale human body according to the human body skeleton data, and constructing a human body multi-scale map by taking body parts as points and relationships among the parts as edges based on the multi-scale human body; a human body motion feature extraction step: introducing the human body multi-scale map into a depth model formed by spatial multi-scale map convolution, and extracting the comprehensive action semantic information of the multi-scale human body; and an action analysis and prediction step: realizing action prediction according tothe comprehensive action semantic information. According to the method, the high-level semantic information of the action can be extracted by utilizing the self-adaptive and dynamic graph structure and the DMGNN, and the action prediction is realized by utilizing the high-level semantic information.

The invention provides a man-machine adversarial action prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: constructing an adversarial space based on man-machine adversarial environment information, and carrying out the modeling to obtain an adversarial space model; analyzing and reasoning the adversarial space model to construct a situation cognition model; performing macroscopic decision based on the situation cognition model, generating a macroscopic strategy plan, and decomposing the macroscopic strategy plan into microoperation behaviors of each action unit; and on the basis of the micro-operation behavior of each action unit and the situation cognition model, performing target selection, path generation, time planning and underlying action generation on each action unit to obtain an action sequence of each action unit. According to the invention, the man-machine adversarial capability of the intelligent agent is improved.

The invention discloses an action prediction method based on a multitask random forest, and the method comprises the following steps: building an action prediction model based on the multitask randomforest through employing a training video marked with an action type label and an observation rate label at the same time; for a newly input video containing incomplete actions, predicting the actiontype of the video by using a multi-task random forest. According to the action prediction method based on the multi-task random forest, aiming at the difficulties that an input video is incomplete andthe observation rate is unknown in action prediction, the performance of an action prediction model is greatly improved by jointly learning classifiers of two tasks, namely action classification andvideo observation rate identification.

The invention discloses a method for predicting acute joint toxicity of three pesticides to photogenic bacteria, which aims to overcome problems that a conventional toxicology acute joint toxicity evaluation technique is large in testing workload and quantitative evaluation and acute joint toxicity action prediction methods are not available. The method for predicting the acute joint toxicity of three pesticides to photogenic bacteria comprises the following steps: (1) performing pretesting, namely confirming the testing concentration of different pesticides in official tests, wherein the step of performing pretesting, namely confirming testing concentrations of different pesticides in official tests comprises the following steps: (1) primarily confirming high, medium and low photogenic bacteria toxicity concentration ranges of a single pesticide; (2) establishing a dosage-effect equation that y is equal to f(x)(x belongs to [C,C']) of the single pesticide; (3) confirming the testing concentrations of different pesticides in the BBD tests; (2) confirming testing schemes through three-factor three-level center combined testing design (BBD); (3) performing official testing, namely, testing the relative light emission inhibition rates of different test groups of photogenic bacteria; and (4) establishing a model to predict the acute joint toxicity of the three pesticides to the photogenic bacteria.