Abstract: The invention discloses a pure end-to-end deep reinforcement learning for training car racing game AI bot that uses only the velocity information extracted from screen for both training and testing phases without using any internal state from game environment, such as the car facing angle. The learned AI bot can play better than the average performance of human players. In addition, the reward function is designed to consist only the velocity value, and use Ape-X distributed training framework combined with a variant of Deep Q Network to solve the sparse training signal problem caused by the reward function of an original design. Moreover, limit learner rate method is designed that improves the training efficiency and training performance. The AI bot trained in this way can achieve performance beyond the average human level and reach a level close to professional players.

Abstract: A method for image analysis of predicted cell metastasis is provided. A host performs principle component analysis (PCA) for analysis and conversion of reference images into hyperspectral image information. Then an image capture unit sends input images to the host. The host converts the input images into hyperspectral images according to the hyperspectral image information and gets spectral information of the hyperspectral images. Next the host selects a plurality of wave bands corresponding to esophageal cancer cells and performs feature computation of the spectral information to generate corresponding features images. Then the host performs convolution of the feature images with kernels to get a convolution result. Later the host matches the convolution result with sample spectra of sample images to get a comparison result. Lastly the host determines whether metastasis of esophageal cancer cells occurs according to the comparison result.

Abstract: A method for correcting abnormal point cloud is disclosed. Firstly, receiving a Primitive Point Cloud Data set by an operation unit for dividing a point cloud array into a plurality of sub-point cloud sets and obtaining a plurality of corresponding distribution feature data according to an original vector data of the Primitive Point Cloud Data set. Furthermore, recognizing the sub-point cloud sets according to the corresponding distribution feature data for correcting recognized abnormal point cloud. Thus, when the point cloud array is rendered to a corresponding image, the color defect of the point cloud array will be improved or decreased for obtaining lossless of the corresponding image.

Abstract: A method for pre-processing geometric model data of a 3D modeling software for deep learning includes steps of: determining a size of a virtual grid that is visualized as a cube based on a smallest value of dimension among values of dimension of objects; generating an empty tensor that is visualized as a cuboid consisting of the virtual grids; assigning an initial value to each of the virtual grids of the empty tensor; replacing the initial value of each of those of the virtual grids with a pre-determined identification attribute value that corresponds uniquely to the property of the corresponding one of the at least one geometric object, so as to generate a 3D geometric model tensor to be used as an input for deep learning; and saving the 3D geometric model tensor in a database in a predefined format.

Abstract: A method for correcting abnormal point cloud is disclosed. Firstly, receiving a Primitive Point Cloud Data set by an operation unit for dividing a point cloud array into a plurality of sub-point cloud sets and obtaining a plurality of corresponding distribution feature data according to an original vector data of the Primitive Point Cloud Data set. Furthermore, recognizing the sub-point cloud sets according to the corresponding distribution feature data for correcting recognized abnormal point cloud. Thus, when the point cloud array is rendered to a corresponding image, the color defect of the point cloud array will be improved or decreased for obtaining lossless of the corresponding image.

Abstract: The invention discloses a pure end-to-end deep reinforcement learning for training car racing game AI bot that uses only the velocity information extracted from screen for both training and testing phases without using any internal state from game environment, such as the car facing angle. The learned AI bot can play better than the average performance of human players. In addition, the reward function is designed to consist only the velocity value, and use Ape-X distributed training framework combined with a variant of Deep Q Network to solve the sparse training signal problem caused by the reward function of an original design. Moreover, limit learner rate method is designed that improves the training efficiency and training performance. The AI bot trained in this way can achieve performance beyond the average human level and reach a level close to professional players.

Abstract: The invention discloses a pure end-to-end deep reinforcement learning for training car racing game AI bot that uses only the velocity information extracted from screen for both training and testing phases without using any internal state from game environment, such as the car facing angle. The learned AI bot can play better than the average performance of human players. In addition, the reward function is designed to consist only the velocity value, and use Ape-X distributed training framework combined with a variant of Deep Q Network to solve the sparse training signal problem caused by the reward function of an original design. Moreover, limit learner rate method is designed that improves the training efficiency and training performance. The AI bot trained in this way can achieve performance beyond the average human level and reach a level close to professional players.

Abstract: The present invention relates to a method for adjusting the strength of turn-based game automatically. The method provides a search algorithm for inquiry and giving decision results. The decision results can be used for filtering and giving filtered results. Then a probability distribution result can be provided to control the virtual node inside the computer host. Furthermore, the game result can be used for adjusting the performance of the virtual node in the game. Thereby, the client node can evaluate its performance in the game conveniently.

Abstract: The present invention relates to a method for adjusting the strength of turn-based game automatically. The method provides a search algorithm for inquiry and giving decision results. The decision results can be used for filtering and giving filtered results. Then a probability distribution result can be provided to control the virtual node inside the computer host. Furthermore, the game result can be used for adjusting the performance of the virtual node in the game. Thereby, the client node can evaluate its performance in the game conveniently.

Abstract: The invention discloses a pure end-to-end deep reinforcement learning for training car racing game AI bot that uses only the velocity information extracted from screen for both training and testing phases without using any internal state from game environment, such as the car facing angle. The learned AI bot can play better than the average performance of human players. In addition, the reward function is designed to consist only the velocity value, and use Ape-X distributed training framework combined with a variant of Deep Q Network to solve the sparse training signal problem caused by the reward function of an original design. Moreover, limit learner rate method is designed that improves the training efficiency and training performance. The AI bot trained in this way can achieve performance beyond the average human level and reach a level close to professional players.

Abstract: The present invention provides a method for scheduling jobs with idle resources. When the computational units of computers are idle, according to the variation of idle time, appropriate jobs may be allocated to the computational units. Then the unscheduled jobs may be completed by using idle time segments. Thereby, the utilization rate of computation resources and the completion rate of jobs may be enhanced.