CROSS-DOMAIN SEGMENTATION FOR ASSIGNING ELECTROMAGNETIC MATERIALS

Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Claim Rejection – 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. Claim(s) 1-7, 9, and 10-21 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yang (US2019150006A1), hereinafter referenced as Yang, in view of the following: Nishikawa (US2021329478A1), hereinafter referenced as Nishikawa, Christensen (Ray Differentials and Multiresolution Geometry Caching for Distribution Ray Tracing in Complex Scenes), hereinafter referenced as Christensen, Chatterjee (Convergence of Gradient Descent For Deep Neural Networks), hereinafter referenced as Chatterjee, Kohli (Semantic Implicit Neural Scene Representations With Semi-Supervised Training) hereinafter referenced as Kohli, and Li (Multi-view Inverse Rendering for Large-scale Real-world Indoor Scenes), hereinafter referenced as Li. Regarding claim 18, Yang teaches a system, comprising: a memory that stores reference radio characteristics; “FIG. 7 is a block diagram of an example computer system 700 ¶ 0123; The computer 702 also includes a database 706 that can hold data for the computer 702 or other component ¶ 0131; one or more processors receive geographic data representing geographic information of a geographic area ¶ 0083” (Yang) and a processor that is connected to the memory, wherein the processor is configured to produce simulated radio characteristics for a three-dimensional scene (3D) by: computing, by a ray tracing method, simulated radio characteristics for the 3D scene based on at least one trainable parameter corresponding to a scene property “geographic information system (GIS) data that includes one or more of a building height map layer, a terrain layer, or a clutter layer ¶ 0014; some layers of the ConvNet contain trainable parameters … For example, the CONV/FC layers perform transformations that are a function of not only the activations in the input volume, but also of trainable parameters (the weights and biases of the neurons) ¶ 0058-0059; A deep neural network can include a convolutional neural network (also referred to as a ConvNet) … the deep neural network 200 receives a GIS map patches 21 ¶ 0068; a deep neural network for received signal strength prediction, both simulated and actual received signal strength data can be used…to generate millions of received signal strength prediction using simulation (e.g., based on the ray tracing method) ¶ 0074; Each of the units can be implemented by one or more processors (e.g., the processor 705) interoperably coupled with one or more computer memory devices (e.g., the memory 707) ¶ 136” (Yang) Yang teaches a processor connected to memory, wherein the processors implementing the deep neural networks ConvNet is configured to predict signal strength (reads on radio characteristic) for the geographic area by: the ConvNet computing, by ray tracing method, simulated signal strength for the geographic area based on at least one trainable parameter corresponding to geographic information input (reads on a scene property). and updating the at least one trainable parameter to minimize a loss function of the simulated radio characteristics and the reference radio characteristics. “some layers of the ConvNet contain trainable parameters (i.e., parameters whose values are update or otherwise derived via the training or learning process) … The trainable parameters in the CONV/FC layers can be trained with gradient descent ¶ 0058; the supervised learning 304 can receive the GIS map patches 302 as the input and the corresponding received signal strength matrices 306 as the ground truth (or a supervisor) and determine the parameters (e.g., the weights and biases of the neurons) of the ConvNet-VOLCANO 308 that minimize a loss function. ¶ 00075; the one or more processors can calculate path loss values corresponding to the predicted received signal strength. ¶ 0110” (Yang) Yang teaches updating the at least one trainable parameter to minimize a loss function of the calculated path loss values corresponding to the predicted received signal strength (reads on simulated radio characteristics) and GIS map patches input (reads on the reference radio characteristics). Yang fails to teach the following: radio characteristics for a three-dimensional scene; a differentiable ray tracer; using gradient-based optimization; and wherein for at least one ray intersection point in the 3D scene, segmentation data associated with the 3D scene is used to estimate the at least one trainable parameter. Nishikawa teaches a processor is configured to produce simulated radio characteristics for a three-dimensional scene. “the computer 100 has a processor 101 ¶ 0171; simulated radio wave strength calculation unit 43 calculates the simulated radio wave strength at each position in the target space for the simulation signal based on the geospatial information on the target space using the theoretical model of radio wave propagation ¶ 0082” (Nishikawa) Nishikawa teaches of a processor which simulated radio wave strength (reads on radio characteristics) for geospatial information (reads on three-dimensional scene). Nishikawa BASE is analogous art with respect to Yang because they are from the same field of endeavor, namely computational models. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang with the feature of Nishikawa to incorporate a processor is configured to produce simulated radio characteristics for a three-dimensional scene. A person of ordinary skill in the art would do such in order to improve computer model accuracies. Yang in view of Nishikawa fail to teach the following: a differentiable ray tracer; using gradient-based optimization; and at least one ray intersection point in the 3D scene, segmentation data associated with the 3D scene is used to estimate the at least one trainable parameter. Christensen teaches computing, by a differentiable ray tracer, the simulated characteristics for the three-dimensional scene “Igehy’s ray differential method11 traces single rays, but keeps track of the difference between each ray and two (real or imaginary) “neighbor” rays. These differences give an indication of the cone/beam size that each ray represents. The curvature at surface intersection points determines how those ray differentials are propagated at specular reflection and refraction.” (Christensen, Page 3, Sec 2.5) PNG media_image1.png 624 1447 media_image1.png Greyscale (Christensen) Christensen computing the simulated characteristics for the 3D scene by a differentiable ray tracer. Christensen BASE is analogous art with respect to Yang in view of Nishikawa because they are from the same field of endeavor, namely computational models. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of Nishikawa with the feature of Christensen to incorporate computing, by a differentiable ray tracer, the simulated characteristics for the three-dimensional scene. A person of ordinary skill in the art would do such in order to improve memory utilization for high computationally intensive graphics. Yang in view of Nishikawa and Christensen fail to teach using gradient-based optimization to minimize a loss function and wherein for at least one ray intersection point in the 3D scene, segmentation data associated with the 3D scene is used to estimate the at least one trainable parameter. Chatterjee teaches updating the at least one trainable parameter using gradient-based optimization to minimize a loss function. “The goal of gradient descent is to find a minimum of a differentiable function… For example, it has been observed that gradient descent can often find global minima of training loss in deep learning [29, 62], which is one of the reasons behind great success of the ‘deep learning revolution’ [14, 40]” ( Page 1-2 Sec 1, Chatterjee) Chatterjee teaches updating the at least one trainable parameter using gradient descent optimization to minimize a loss function via finding global minima of training loss. Chatterjee BASE is analogous art with respect to Yang in view of Nishikawa and Christensen because they are from the same field of endeavor, namely computational models. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of Nishikawa and Christensen with the feature of Chatterjee to incorporate updating the at least one trainable parameter using gradient descent optimization to minimize a loss function via finding global minima of training loss. A person of ordinary skill in the art would do such in order to improve optimize learning models. Yang in view of Nishikawa, Christensen, and Chatterjee fail to teach wherein for at least one ray intersection point in the 3D scene, segmentation data associated with the 3D scene is used to estimate the at least one trainable parameter. Kohli teaches wherein for at least one ray intersection point in the 3D scene, segmentation data associated with the 3D scene is used to estimate the at least one trainable parameter. “a scene is modeled as a function that maps world coordinates x to a feature representation of local scene properties v: PNG media_image2.png 110 1312 media_image2.png Greyscale Images are synthesized from this 3D representation via a differentiable neural renderer consisting of two parts. The first is a differentiable ray marcher which finds intersections of camera rays with scene geometry by marching along a ray away from a camera … A hypernetwork [25] HN maps embedding vectors zj to the weights wj of the respective scene representation network: PNG media_image3.png 106 1326 media_image3.png Greyscale … z is optimized to obtain a new scene embedding via minimizing image reconstruction error. (Page 3-4 Section 3.1); We formalize dense 3D semantic segmentation as a function that maps a world coordinate x to a distribution over semantic labels y … we define the Segmentation Renderer SEG, a function that maps a feature vector v to a distribution over class labels y: PNG media_image4.png 110 1351 media_image4.png Greyscale Since v is a function of x, we may enforce a per-pixel cross-entropy loss on the SEG output at any world coordinate x: PNG media_image5.png 201 1185 media_image5.png Greyscale In any of these cases, a new code vector is inferred by freezing all network weights, initializing a new code vector z, and optimizing z to minimize image reconstruction and/or cross entropy losses (Page 4 Sec 3.2)” (Kohli) x is a point along the ray which is an intersection point in the 3D representation (reads on as least one ray intersection point in the 3D scene). The feature vector v is a function of x via SRN, see equation 1. SRN encodes w weights via HN, in which hypernetwork HN defines w using z. w is trained by adjusting z, see equation 2. (w reads on at least trainable parameter). SEG maps the feature vector v to a distribution over class labels to determine the predicted class probabilities outputted as y, see equation 3 (y reads on segmentation data). y is used to compute per-pixel cross-entropy loss, see equation 4. Cross entropy loss function is then used to optimize z to train w. Kohli teaches of at least one ray intersection point in the 3D scene x, segmentation data y associated with the 3D scene is used to estimate the at least one trainable parameter w. Kohli BASE is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, and Chatterjee because they are from the same field of endeavor, namely computational models. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, and Chatterjee with the feature of Kohli to use at least one ray intersection point in the 3D scene x, segmentation data y associated with the 3D scene is used to estimate the at least one trainable parameter w. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Yang in view of the following: Nishikawa, Christensen, Chatterjee and Kohli does not explicitly teach wherein the loss function minimizes variance of one or more properties corresponding to radio materials within each segmentation class. But Li does. Li teaches: wherein the loss function minimizes variance of one or more properties corresponding to “We address this problem by leveraging the priors of semantics and room segmentation. The semantic images are predicted by learning-based models [11] and room segmentation is calculated by the occupancy grid [15]. Our segmentation-based strategy has three phases. Details of each stage are described in the following subsections…Stage I: albedo initialization… we apply a semantic smoothness constraint to encourage that the albedo/roughness/feature is as close to the mean within the class as possible… Stage II: VHL-based sampling and semantics-based propagation… by leveraging the prior of semantic segmentation, the reasonable roughness of highlight regions would be propagated into the regions with same semantics… Stage III: Segmentation-based fine-tuning. In the last phase, we fine-tune all material textures based on the priors of semantic segmentation and room segmentation” (Li, Section 3.3) PNG media_image6.png 118 728 media_image6.png Greyscale (Li, Equation 4); Li teaches semantic smoothness constraint applied in the segmentation-based material optimization, see equation 4 (reads on loss function). The constraint encourages each value to be as close to the mean within the class as possible (reads on function minimizes variance of one or more properties because every value is being pulled and minimized to close to the mean). The constrained values are the albedo, roughness, and feature of the material textures corresponding to each class of the semantic images predicted by the learning -based models (reads one or more properties corresponding to materials within each segmentation class). Li is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li because they are from the same field of endeavor, namely computational models. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Li to apply semantic smoothness constraint applied in the segmentation-based material optimization to radio material. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Claim 1 is rejected using the same rationale or bases as applied to claim 18. Claim 23 is rejected using the same rationale or bases as applied to claim 18. Regarding claim 2, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teach the computer-implemented method of claim 1, and additionally teaches the following. Kohli teaches wherein the segmentation data comprises an image-based segmentation map, light detection and ranging (LIDAR) based segmentation map, point-cloud segmentation data, surface segmentation map, or a volumetric segmentation map used to obtain the at least one trainable parameter associated with object classes identified in the segmentation mask, and further comprising translating a two-dimensional segmentation data corresponding to the 3D scene to produce the segmentation data. “Here we view the recently proposed scene representation networks (SRNs) from a representation learning perspective in order to infer multi-modal, compact 3D representations of objects from 2D images. We take the latent 3D feature representation of SRNs, learned in an unsupervised manner given only posed 2D RGB images, and map them to a set of labeled semantic segmentation maps (Page 2 Sec 1); This training is supervised with human-labeled, posed semantic segmentation masks of a small subset of the training images (Page 5, Sec 3.3); We formalize dense 3D semantic segmentation as a function that maps a world coordinate x to a distribution over semantic labels y. This can be seen as a generalization of point cloud- and voxel-grid-based semantic segmentation approaches (Page 4 Sec 3.2)” (Kohli) Kohli teaches of point cloud- and voxel-grid-based semantic segmentation approaches (reads on point-cloud segmentation data) wherein 2D RGB images are mapped to labeled semantic segmentation maps (reads on two-dimensional segmentation mask). These semantic segmentation masks are used as training inputs (one trainable parameter associated with object classes identified in the segmentation mask) and translated into 3D segmentation data via the SRN feature representation (reads on translating a two-dimensional segmentation mask corresponding to the 3D scene to produce the segmentation data). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate point cloud- and voxel-grid-based semantic segmentation approaches wherein posed 2D RGB images are mapped to labeled semantic segmentation masks for training input and translated into 3D segmentation data via the SRN feature representation. A person of ordinary skill in the art would do such in order to improve computational models. Claim 19 is rejected using the same rationale or bases as applied to claim 2. Regarding claim 3, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teach the computer-implemented method of claim 2, and additionally teaches the following. Kohli teaches wherein the two-dimensional segmentation mask comprises probability distributions of object class identifiers or the object class identifiers. PNG media_image5.png 201 1185 media_image5.png Greyscale “yj is a one-hot ground-truth class label with c number of classes” (Kohli, Page 4 Sec 3.2) yj is the true mapping of segmentation masks and class labels values (reads on two-dimensional segmentation masks) corresponds to j to c class labels (reads on object class identifiers) according to the true distribution with the structure a cross-entropy loss function, see equation 4 (reads on the segmentation mask comprise of object class identifiers). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to wherein the segmentation mask yj comprise of object class identifiers. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Regarding claim 4, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Kohli teaches wherein the segmentation data are included in the at least one trainable parameter. “For an object with c semantic classes, the optimization parameters are matrix W ∈ R256xc and bias b ∈ Rc. (Page 12 Supplemental Sec 1); The key idea of SRNs is to encode a scene in the weights w ∈ Rl of a fully connected neural network, the SRN itself … To this end, a scene is modeled as a function that maps world coordinates x to a feature representation of local scene properties v: PNG media_image2.png 110 1312 media_image2.png Greyscale … A hypernetwork [25] HN maps embedding vectors zj to the weights wj of the respective scene representation network: PNG media_image3.png 106 1326 media_image3.png Greyscale … z is optimized to obtain a new scene embedding via minimizing image reconstruction error. Segmentation Renderer SEG, a function that maps a feature vector v to a distribution over class labels y: PNG media_image4.png 110 1351 media_image4.png Greyscale Page 3-4 Section 3.1-3.2” (Kohli) SEG maps the feature vector to a distribution over class labels to determine the predicted class probabilities outputted as y, see equation 3 (y reads on segmentation data). The feature vector v inputted into SEG is derived from SRN which encodes w weights. Hypernetwork HN defines w using z, see equation 2. w is trained by adjusting z. (w reads on at least one trainable parameter). Kohli teaches of segmentation data y that includes at least one trainable parameter w because v = SRN (x; w) and y = SEG(v), the output of y is a function of w through v. w are used in generating the vector feature v, which is then inputted into SEG to produce the segmentation data y. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate the segmentation data, y, to include in the at least one trainable parameter, w. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Regarding claim 5, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Kohli teaches wherein a neural component estimates the at least one trainable parameter based on the at least one ray intersection point and the segmentation data associated with the at least one ray intersection point. “a scene is modeled as a function that maps world coordinates x to a feature representation of local scene properties v: PNG media_image2.png 110 1312 media_image2.png Greyscale The key idea of SRNs is to encode a scene in the weights w ∈ Rl of a fully connected neural network, the SRN itself … A hypernetwork [25] HN maps embedding vectors zj to the weights wj of the respective scene representation network: PNG media_image3.png 106 1326 media_image3.png Greyscale … We formalize dense 3D semantic segmentation as a function that maps a world coordinate x to a distribution over semantic labels y … we define the Segmentation Renderer SEG, a function that maps a feature vector v to a distribution over class labels y: PNG media_image4.png 110 1351 media_image4.png Greyscale … Since v is a function of x, we may enforce a per-pixel cross-entropy loss on the SEG output at any world coordinate x: PNG media_image5.png 201 1185 media_image5.png Greyscale … In any of these cases, a new code vector is inferred by freezing all network weights, initializing a new code vector z, and optimizing z to minimize image reconstruction and/or cross entropy losses (Page 4 Sec 3.2)” (Kohli) Hypernetwork estimates trainable parameter w by optimizing z. Cross entropy losses is used to optimize z. Cross-entropy loss, see equation 4, is calculated using SEG. SEG maps the feature vector v to a distribution over class labels to determine the predicted class probabilities to get y, see equation 3 (y reads on segmentation data). The feature vector v is a function of x via SRN, see equation 1, which x is a point along the ray which is an intersection point in the 3D representation. (reads as least one ray intersection point in the 3D scene). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate hypernetwork estimates trainable parameter w based on the at least one ray intersection point x and the segmentation data y associated with the at least one ray intersection point x. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Regarding claim 6, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 5, and additionally teaches the following. Kohli teaches wherein the at least one ray intersection point is encoded into a higher dimensional space for input to the neural component. “The key idea of SRNs is to encode a scene in the weights w ∈ Rl of a fully connected neural network, the weights w ∈ Rl of a fully connected neural network, the SRN itself. To this end, a scene is modeled as a function that maps world coordinates x to a feature representation of local scene properties v (Page 3-4 Sec 3.1); PNG media_image2.png 110 1312 media_image2.png Greyscale … we take the pre-trained features v ∈ R256 from the neural scene representation and use a simple linear transformation to map those features to class probabilities for each pixel. (Page 12 Supplemental Sec 1) The feature vector v is a function of x via SRN, see equation 1. The SRN maps a 3D world coordinate x to a n-dimensional feature vector v, lifting from R3 to Rn. This is a higher dimensional encoding of the ray intersection point before it is inputted into the segmentation renderer neural component. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate at least one ray intersection point is encoded into a higher dimensional space for input inputted into the segmentation renderer neural component. A person of ordinary skill in the art would do such in order to improve computational model accuracies. Regarding claim 7, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 5, and additionally teaches the following. Kohli teaches wherein the updating adjusts weights that are applied, by the neural component, to the at least one ray intersection point and the segmentation data to estimate the at least one trainable parameter. “We train our models using ADAM with a learning rate of 4e−4… we take the pre-trained features v ∈ R256 from the neural scene representation and use a simple linear transformation to map those features to class probabilities for each pixel…Image reconstruction loss and cross-entropy loss are weighted 200 : 8, such that their magnitudes are approximately equal. (Page 12 Supplemental Sec 1); Since v is a function of x, we may enforce a per-pixel cross-entropy loss on the SEG output at any world coordinate x: PNG media_image5.png 201 1185 media_image5.png Greyscale … We can now train the segmentation renderer end-to-end composed with the same architecture used to pre-train the scene representation. In any of these cases, a new code vector is inferred by freezing all network weights, initializing a new code vector z, and optimizing z to minimize image reconstruction and/or cross entropy losses (Page 4 Sec 3.2)” (Kohli) SEG(v) or y (reads on segmentation data) is used to compute per-pixel cross-entropy loss, see equation 4. Cross entropy loss function is then used to optimize z to train weight w see equation 2. (reads on the segmentation data estimating the at least one trainable parameter). Moreover, the training process adjusts the weights of the linear segmentation render via the ADAM optimizer (ADAM optimizer reads on neural component). These wights are applied to the feature vectors v, which are derived from ray intersection points see equation 1 and 2. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Kohli with the feature of Kohli to updating adjusts weights that are applied, by the ADAM optimizer, to the at least one vector v, which is derived from ray intersection points and the segmentation data, to estimate the at least one trainable parameter w. A person of ordinary skill in the art would do such in order to improve computational model accuracy. Regarding claim 9, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1 and additionally teaches the following. Kohli teaches: A loss function based on the segmentation data. we define the Segmentation Renderer SEG, a function that maps a feature vector v to a distribution over class labels y: PNG media_image4.png 110 1351 media_image4.png Greyscale Since v is a function of x, we may enforce a per-pixel cross-entropy loss on the SEG output at any world coordinate x: PNG media_image5.png 201 1185 media_image5.png Greyscale In any of these cases, a new code vector is inferred by freezing all network weights, initializing a new code vector z, and optimizing z to minimize image reconstruction and/or cross entropy losses (Page 4 Sec 3.2)” (Kohli) SEG maps the feature vector v to a distribution over class labels to determine the predicted class probabilities outputted as y, see equation 3 (y reads on segmentation data). y is used to compute per-pixel cross-entropy loss, see equation 4. Kohli teaches of a loss function based on the segmentation data. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate a loss function based on segmentation data. A person of ordinary skill in the art would do such in order to improve computational accuracies. Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li fail to teach wherein the loss function includes a regularization term based on the segmentation data. Nishikawa does. Nishikawa teaches wherein the one or more properties comprise electromagnetic properties predicted by a model at points assigned to each segmentation class. “a simulated radio wave strength calculation unit 33 configured to calculate a simulated radio wave strength at each position in a target space for a simulation signal based on geospatial information on the target space” (Nishikawa, Abstract) “Here, the prediction model is a model that expresses the radio wave strength at any position (which is the target position here) in the target space using the radio wave strengths at the installation positions of the first observation apparatuses 10. That is, when any position is input, the prediction model in the “learning mode” outputs the simulated radio wave strength at the input position using the simulated radio wave strengths at the installation positions of the first observation apparatuses 10 calculated by the simulated radio wave strength calculation unit 43.” (Nishikawa, ¶ 87); “Then, the simulated radio wave strength calculation unit 43 partitions the quadrilateral area S in a lattice form, and defines each point on the lattice as a point where the first observation apparatuses 10 or the second observation apparatus 20 may be arranged on the simulation model. Then, the simulated radio wave strength calculation unit 43 sets the position of the signal emission source (also referred to as the transmission position) from among the points included in the quadrilateral area S based on the position information on the signal emission source included in the geospatial information.” (Nishikawa, ¶ 101); Nishikawa teaches a prediction model that simulates radio waves strength at any given position in the target space is inputted (reads on the one or more properties comprise electromagnetic properties predicted by a model at points) wherein the simulated radio wave strength calculation are arranged on the simulation model and position information (reads on points assigned to each segmentation class). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Nishikawa to incorporate a prediction model that simulates radio waves strength at any given position in the target space is inputted wherein the simulated radio wave strength calculation are arranged on the simulation model and position information. A person of ordinary skill in the art would do such in order to improve computational models’ accuracies. Regarding claim 10, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teach the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein the at least one trainable parameter includes one or more of the segmentation data, scene geometry, configuration of reconfigurable intelligent surfaces and meta materials, antenna patterns, array geometries, and transmitter and receiver orientations and positions. “Supervised learning 304 can be performed to generate a trained deep neural network, ConvNet-VOLCANO 308, based on the simulated training data obtained from the VOLCANO simulator 305 . For example, the supervised learning 304 can receive the GIS map patches 302 as the input. ¶ 0075; The horizontal LOS angle θ 1210 can refer to the horizontal angle θ (in the x-y plane) of an LOS path 1240 from the antenna 1235 of the base station 1230 to the cube 1220 … A vertical LOS angle tensor can include vertical LOS angles at different locations in a geographic area with respect to a base station in the geographic area. ¶ 0095-0096; the one or more processors use a multi-dimensional tensor (e.g., the above-described 4D tensor that includes five channels of 3D tensors) based on the geographic data (and the antenna and power information of the base station into as the input to the convolutional neural network. ¶ 0106” (Yang) Yang teaches wherein the at least one trainable parameter includes one or more GIS map patches (GIS inputs encode structured spatial information regarding the environment reads on segmentation data and scene geometry), antenna and transmit power information, spatial tensors representing line of sight angles, input data corresponding to environmental geometry antenna configurations, and transmitter and receiver spatial orientation and positions. Regarding claim 11, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teach the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein the scene property comprises at least one of relative permittivity, conductivity, effective roughness, and permeability of object surfaces and scattering functions. “The geographic data can include geographic image data such as geographic information system (GIS) data (e.g., the GIS map patches 110 , 210 , 302 , or 402 ) that includes one or more of a building height map layer, a terrain layer, or a clutter layer representing geographic information of a geographic area. ¶ 0083; the one or more processors further compute or otherwise determine a multi-dimensional … based on the geographic data…and the antenna and power information of the base station, for example, based on known physical laws (e.g., free-space path loss and signal attenuation based on material permittivity). ¶ 0086” (Yang) Yang teaches the scene property comprises geographic data, antenna, and power information of the base station based on known physical laws of electromagnetic properties of the surfaces (reads on permittivity, conductivity, effective roughness, and permeability of object surfaces and scattering functions). Regarding claim 12, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein the simulated radio characteristics comprise one or more of channel impulse responses, channel frequency responses, path delays, path losses, angles of arrival, angles of departure, amplitudes, powers, delay spread, Doppler spread, angular spread, power-delay-angular profile, and a number of paths. “The simulator can be, for example, a commercial received signal strength simulator such as VOLCANO simulator 305 , or any other simulators that implement one or more of a statistical method, a deterministic method, or a combination of these and other methods for received signal strength prediction. ¶ 0117; the simulated received signal strength representing simulated signal strength of wireless signals received at different locations in the second geographic area ¶ 0119” Yang teaches wherein the simulated received signal strength (reads on simulated radio characteristics) comprise of signal strength of wireless signals received at different locations (reads on powers). Regarding claim 13, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein the simulated radio characteristics for the 3D scene are computed based on configured parameters. “a ConvNet takes GIS data as an input, and predicts received signal strength matrix as an output… A ConvNet can have hundreds of millions of parameters (or weights in neural network terminology) that are learned from training ¶ 59” (Yang) Yang teaches received signal strength (reads on radio characteristics) are computed based on a ConvNet can have hundreds of millions of parameters which can be earned and trained (reads on configured parameters) Nishikawa teaches the simulated radio characteristics for the 3D scene “simulated radio wave strength calculation unit 43 calculates the simulated radio wave strength at each position in the target space for the simulation signal based on the geospatial information on the target space using the theoretical model of radio wave propagation ¶ 0082(Nishikawa) Nishikawa teaches of simulated radio wave strength (reads on radio characteristics) for geospatial information (reads on three-dimensional scene). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Nishikawa to incorporate a simulated radio wave for geospatial information. A person of ordinary skill in the art would do such in order to improve computational model accuracies and precisions. Regarding claim 14, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches: wherein at least one of the steps of computing or updating is performed on a server or in a data center and the simulated radio characteristics or an image generated from the simulated radio characteristics is streamed to a user device. “computer 702 is intended to encompass any computing device such as a server, desktop ¶ 0125; The computer 702 can receive requests over network 730 from a client application … and respond to the received requests by processing the received requests using an appropriate software application(s). In addition, requests may also be sent to the computer 702 from internal users ¶ 0127; a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the subject matter described in this specification ¶ 0175; The computing system can include clients and servers …The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. ¶ 0176” Yang teaches of wherein at least one of the steps of computing or updating is performed on a server and the simulated radio characteristics or an image generated from the simulated radio characteristics is streamed to a user device via a client server relationship. Regarding claim 15, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches: wherein at least one of the steps of computing or updating is performed within a cloud computing environment. “In some implementations, one or more components of the computer 702 may be configured to operate within environments, including cloud-computing-based, local, global, or other environment (or a combination of environments). ¶ 0125” (Yang) Yang teaches wherein at least one of the steps of computing or updating within cloud-computing-based environments. Regarding claim 16, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein at least one of the steps of computing or updating is performed for training, testing, or certifying a neural network employed in a machine, robot, or autonomous vehicle. “method 500 for predicting received signal strength in a telecommunication network using a deep neural network…The method 500 can be implemented by one or more processors that execute a convolutional neural network that includes a number of convolution layers ¶ 0078; The convolutional neural network can be a deep neural network (e.g., the deep neural network 120 or 200 , the ConvNet-VOLCANO 308 , and ConvNet-Real 408). ¶ 0079” (Yang) Yang teaches wherein at least one of the steps of computing or updating is performed for training, testing, or certifying a neural network employed in processor(s). Regarding claim 17, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Yang teaches wherein at least one of the steps of computing or updating is performed on a virtual machine comprising a portion of a graphics processing unit. “the described techniques can be 10,000 times faster in computation compared to ray tracing methods using normal workstations with a graphic processing unit (GPU). ¶ 64; the illustrated computer 702 is intended to encompass any computing device such as … any other suitable processing device, including physical or virtual instances (or both) of the computing device ¶ 124; Yang teaches wherein at least one of the steps of computing or updating is performed on a virtual instances comprising a portion of a graphics processing unit. Claim 20 is rejected using the same rationale or bases as applied to claim 18 and the mentioned structure. Additionally, claim 20 recites the following structure: A non-transitory computer-readable media storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform the steps as taught by Yang. “a non-transitory computer-readable media storing computer instructions for predicting received signal strength in a telecommunication network, that when executed by one or more processors, cause the one or more processors to perform the steps of ¶ 0008” (Yang); Regarding claim 21, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches non-transitory computer-readable media of claim 20, and additionally teaches the following. Yang teaches wherein the ray tracer computes paths of electromagnetic waves. “In some implementations, the described techniques can improve accuracy of received signal strength prediction. For example, the described techniques using the deep neural network can capture complex relations between the GIS data and the received signal strength, whereas the statistical models, such as HATA, has only a few parameters and can only capture simple relations between the GIS data and the received signal strength. In some implementations, the described techniques can obtain more accurate results than deterministic methods such as ray tracing methods, for example, in the case where the input is incomplete GIS data.” Yang teaches of ray tracer computes electromagnetic waves, received signal strength. Christensen teaches differentiable ray tracer computes paths. “Igehy’s ray differential method11 traces single rays, but keeps track of the difference between each ray and two (real or imaginary) “neighbor” rays. These differences give an indication of the cone/beam size that each ray represents. The curvature at surface intersection points determines how those ray differentials are propagated at specular reflection and refraction.” (Christensen, Page 3, Sec 2.5) PNG media_image1.png 624 1447 media_image1.png Greyscale (Christensen) Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Christensen to incorporate the differentiable ray tracer computes paths. A person of ordinary skill in the art would do such in order to improve memory utilization for high computationally intensive graphics. Claim(s) 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Azinovic (Inverse Path Tracing for Joint Material and Lighting Estimation), hereinafter referenced as Azinovic. Regarding claim 8, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches the computer-implemented method of claim 1, and additionally teaches the following. Kohli teaches wherein all ray intersection points associated with a given object class identifier, according to the segmentation data, have association for at least one trainable parameter. “a scene is modeled as a function that maps world coordinates x to a feature representation of local scene properties v: PNG media_image2.png 110 1312 media_image2.png Greyscale …The key idea of SRNs is to encode a scene in the weights w ∈ Rl of a fully connected neural network, the SRN itself … A hypernetwork [25] HN maps embedding vectors zj to the weights wj of the respective scene representation network: PNG media_image3.png 106 1326 media_image3.png Greyscale … We formalize dense 3D semantic segmentation as a function that maps a world coordinate x to a distribution over semantic labels y … we define the Segmentation Renderer SEG, a function that maps a feature vector v to a distribution over class labels y: PNG media_image4.png 110 1351 media_image4.png Greyscale (Page 4 Sec 3.2)” (Kohli) x is a point along the ray which is an intersection point in the 3D representation (ray intersection point). x is used to determine the feature vector v through SRN, see equation 1. SRN which encodes w weights (reads all ray intersection points having association with trainable parameters). Segmentation Renderer SEG, maps a feature vector to a distribution over class labels to produce y (reads on ray intersection points associated with a given object class identifier in accordance with segmentation data). coordinates for the ray intersection points are “these implicit representations encode a scene in the weights of a neural network which can be queried at any coordinate to produce these same scene properties.” (Kholi, Section Abstract); “At every step, it queries SRN at the current world coordinates and translates the resulting feature vector into a step length.” (Kholi, Section 3.1); “We formalize dense 3D semantic segmentation as a function that maps a world coordinate x to a distribution over semantic labels y. This can be seen as a generalization of point cloud- and voxel-grid-based semantic segmentation approaches [11, 56, 58], which label a discrete set of world coordinates, sparsely sampling an underlying, continuous function.” (Kholi, Section 3.2); Kholi queries the points of the coordinates for the ray intersection points to get the shared scene properties (reads on coordinates for the ray intersection points are used to obtain the shared value) Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Kohli to incorporate all ray intersection points x associated with a given class label, according to the segmentation data y, have association for the at least one trainable parameter w, and queries the points of the coordinates for the ray intersection points to get the shared scene properties. A person of ordinary skill in the art would do such in order to improve computational efficiency. Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li fail to teach all ray intersection are associated with a shared value for the at least one trainable parameter and coordinates for the ray intersection points are used to obtain the given object class identifier that is used to obtain the shared value. Azinovic teaches all ray intersection are associated with a shared value for the at least one trainable parameter and “At every interaction vertex xi of the light path, there is a bidirectional reflectance distribution function (BRDF)” (Azinovic Section 3); “The optimization is performed on a per-object basis, i.e., every object has a single unknown emission and a set of material parameters that are assumed constant across the whole object. We show that this is enough to obtain accurate lighting and an average constant value for the albedo of an object.” (Azinovic, Section 4.2); “Figure 11: Evaluation on real scenes: (right) input is 3D scanned geometry and photographs. We employ object instance segmentation to estimate the emission and the average albedo of every object in the scene” (Azinovic, Figure 11) Azinovic teaches a bidirectional reflectance distribution function at every interaction vertex of the light path (all ray intersection), with each object are identified by their object instance segmentation (reads on object class identifier). Each segmented object is assigned a single set of unknown material parameters which is held constant across the whole object (reads on all ray intersection are associated with a shared value for at least one trainable parameter. That is because every vertex of that object uses the same optimized set of unknown), and that set is selected by the object based on which label the point belongs to (reads coordinates for the ray intersection points are used to obtain the given object class identifier that is used to obtain the shared value. That is because the object label determines which parameter set is applied). Azinovic BASE is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li because they are from the same field of endeavor, namely computational modeling. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Azinovic to incorporate a bidirectional reflectance distribution function at every interaction vertex of the light path wherein each segmented object is assigned a single set of unknown material parameters which is held constant across the whole object and that set is selected by the object based on which label the point belongs to. A person of ordinary skill in the art would do such in order to improve computational efficiency. Claim(s) 22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Ferrari (US11742965B2) hereinafter referenced as Ferrari, and Gormley (US20130176973A1) hereinafter referenced as Gormley. Claim 22 is rejected using the same rationale or bases as applied to claim 18 and the mentioned structure. Additionally, regarding claim 22, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li teaches adjusting, based on the simulated signal characteristics, an antenna position, an antenna direction, or directional signal transmission for an antenna array of one or more wifi access points to optimize signal coverage “The predicted path loss values can represent propagation attenuation of the wireless signals at the different locations in the geographic area. Automatic site planning can include, for example, improving, optimizing, or otherwise adjusting antenna configurations (e.g., the antenna locations, antenna beam directions, transmitting power, and in some cases antenna radiation pattern) or other parameters of a base station in the geographic area based on the predicted received signal strength. ¶ 0110” (Yang) Yang teaches of adjusting antenna configurations (reads on an antenna position) for a base station (reads on antenna array of the wifi access points) based on predicted path loss values (simulated signal characteristics). However, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li fail to teach improving wifi performance in a residential or commercial environment and wifi signal characteristics; Ferrari does. Ferrari teaches improving wifi performance in a residential or commercial environment and wifi signal characteristics; “A visualization of the wireless signal propagation can help understanding the signal propagation (i.e., assessing the signal propagation behavior) and validating the signal propagation based on signal level measurements from APs and sensors so that an optimized wireless network can be designed as to where to place or how to configure Wi-Fi APs. ¶ 0005; FIG. 5 illustrates an example 3-D visualization 300 of Wi-Fi AP RF signal propagation. In the 3-D visualization 300, the 3-D visualization of a building plan (e.g., floor plan) is overlaid with RF propagation patterns. ¶ 0069” (Ferrari). PNG media_image7.png 562 936 media_image7.png Greyscale Ferrari teaches optimizing wireless network (reads on improving wifi performance) moreover Wi-Fi access points (reads on wifi signal characteristics) in building plans of commercial building, see figure 5 (reads on commercial environment). Ferrari BASE is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li because they are from the same field of endeavor, namely computational modeling. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, and Li with the feature of Ferrari to optimizing wireless moreover Wi-Fi access points in building plans of commercial building. A person of ordinary skill in the art would do such in order to improve computational usage and efficacy. Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li and Ferrari fail to teach the following: wherein a focused beam of radio waves is created by adjusting the directional signal transmission for the antenna array using a beamforming technique to manipulate the phase and amplitude of the signals at each antenna element in the antenna array. But Gormley does. Gormley teaches: wherein a focused beam of radio waves is created by adjusting the directional signal transmission for the antenna array using a beamforming technique to manipulate the phase and amplitude of the signals at each antenna element in the antenna array. “Wireless communications systems can use transmit beamforming to improve the level of the signal seen at a desired receiver, as well as reduce the levels of interference seen at other receivers. The interference reduction capability of beamforming can be particularly advantageous in cellular wireless systems, where high levels of interference can severely reduce the capacity of such systems. ¶ 2; Beamforming requires at least two transmit or receive sensors. ¶ 3; An additional possibility with beamforming is the ability to reduce interference by choosing phases and signal amplitudes that can cause the signals received at or transmitted by a different mobile station to cancel. ¶ 4; FIGS. 2-5 show some examples of how the phase and amplitude differences between the two signals 104 and 106 arriving at a mobile device (e.g., 108), or a customer premise equipment (CPE), can impact the combined signal that the receiver sees. A beamforming system (e.g., 100) can control the relative amplitudes and phases at the transmitter so that the combined signal seen at the receiver (e.g., 108) can have increased amplitude, or can have reduced amplitude. ¶ 8” (Gormley); Gormley teaches a beamforming system which forms a beam wherein by reducing interference by choosing phases and signal amplitudes that can cause the signals received at or transmitted by a different mobile station to cancel using at least two transmit or receive sensors (reads on a focused beam of radio waves is created by adjusting the directional signal transmission for the antenna array). Moreover, the beamforming system can control the relative amplitudes and phases at the transmitter so that the combined signal seen at the receiver can have increased amplitude, or can have reduced amplitude (reads on using a beamforming technique to manipulate the phase and amplitude of the signals at each antenna element in the antenna array). Gormley BASE is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, and Ferrari because they are from the same field of endeavor, namely signal transmission and simulation. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, and Ferrari with the feature of Gormley to incorporate a beamforming system which forms a beam wherein by reducing interference by choosing phases and signal amplitudes that can cause the signals received at or transmitted by a different mobile station to cancel using at least two transmit or receive sensors. A person of ordinary skill in the art would do such in order to improve cellular wireless systems efficacy. Claim(s) 24 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Ferrari, Gormley, and Desclos (US11742965B2) hereinafter referenced as Desclos. Regarding claim 24, Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, and Ferrari teach the computer-implemented method of claim 22, and additionally Gormley teaches: wherein positions of the elements of the antenna array are unchanged while the direction of the focused beam of radio waves is adjusted towards wifi users to accomplish at least one of enhancing capacity of a wifi signal, enhancing coverage of a wifi signal, increasing data rates, extending range, or reducing interference. “Wireless communications systems can use transmit beamforming to improve the level of the signal seen at a desired receiver, as well as reduce the levels of interference seen at other receivers. The interference reduction capability of beamforming can be particularly advantageous in cellular wireless systems, where high levels of interference can severely reduce the capacity of such systems. ¶ 2; Beamforming requires at least two transmit or receive sensors. ¶ 3; An additional possibility with beamforming is the ability to reduce interference by choosing phases and signal amplitudes that can cause the signals received at or transmitted by a different mobile station to cancel. ¶ 4; FIGS. 2-5 show some examples of how the phase and amplitude differences between the two signals 104 and 106 arriving at a mobile device (e.g., 108), or a customer premise equipment (CPE), can impact the combined signal that the receiver sees. A beamforming system (e.g., 100) can control the relative amplitudes and phases at the transmitter so that the combined signal seen at the receiver (e.g., 108) can have increased amplitude, or can have reduced amplitude. ¶ 8” (Gormley); Gormley teaches a beamforming system that controls the relative amplitudes and phases at the transmitter (reads on positions of the elements of the antenna array are unchanged because the beam is shaped by adjusting the phase and amplitude of the signals fed to the sensors rather than by moving the sensors) so that the combined signal at the receiver has increased or reduced amplitude as to reduce interference (reads on the direction of the focused beam of radio waves is adjusted towards wifi users to accomplish reducing interference). Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, Ferrari with the feature of Gormley to incorporate a beamforming system that controls the relative amplitudes and phases at the transmitter so that the combined signal at the receiver has increased or reduced amplitude as to reduce interference. A person of ordinary skill in the art would do such in order to improve cellular wireless systems efficacy and reduce interference. Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, Ferrari, and Gormley fail to explicitly teach positions of the elements of the antenna array are unchanged. But Desclos does. Desclos teaches: positions of the elements of the antenna array are unchanged “Radiation mode selection for 2.4 GHz and 5 GHz antennas integrated into an external enclosure on a WLAN access point allows for independent optimization of the antenna systems for the two frequency bands without requiring antenna movement or positioning.” (Desclos, Abstract); Desclos teaches of an antenna system wherein the antenna movement or positioning is not required (reads on positions of the elements of the antenna array are unchanged). Desclos is analogous art with respect to Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, Ferrari, and Gormley because they are from the same field of endeavor, namely signal transmission and simulation. Before the effective filling date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify Yang in view of the following: Nishikawa, Christensen, Chatterjee, Kohli, Li, Ferrari, and Gormley with the feature of Desclos to incorporate an antenna system wherein the antenna movement or positioning is not required. A person of ordinary skill in the art would do such in order to increase in the number of available radiation patterns or modes, and the capability to dynamically optimize a multi-external antenna system for specific environments are reduced. Applicant Argument Applicant argues that no reference teaches a loss of simulated radio characteristics that minizines variances of radio material properties within each segmentation class in regarding claims 1, 18, 20. Applicant argues Kohli’s cross-entropy loss function is the loss function claimed. Applicant argues Nair does not teach radio characteristics, radio materials, segmentation classes, or within-class variance. Moreover, in regard to claim 9, Applicant argues that Nair does not teach electromagnetic properties predicted by a model at point assigned to each segmentation. Applicant argues Ravanbakhsh’s parameter sharing concerns tied network weights and not class based shared scene parameters, thus Ravanbakhsh does not teach coordinate used to obtain a class identifier in regard to claim 8. Applicant argues the combination does not teach a focused beam created by adjusting phase and amplitude at each antenna element. Applicant has added claim 22 accordingly. Applicant argues hindsight for 103 rejection combination. Response to Argument Applicant’s argument that the cited references do not teach a loss of simulated radio characteristics that minizines variances of radio material properties within each segmentation class in regarding claims 1, 18, 20 is moot in view of new grounds of rejection. Li (Multi-view Inverse Rendering for Large-scale Real-world Indoor Scenes) teaches a semantic smoothness constraint that minizines variances of material properties within each segmentation class. Yang and Nishikawa supply the teachings for radio material properties. Additionally, Kohli’s cross-entropy is longer cited for the loss function claimed. Kohli is relied upon for world coordinate to semantic labels now. Applicant’s argument regarding Nair is moot because Nair is no longer relied on. The rejection now relies on Nishikawa’s teachings of the electromagnetic properties predicted by a model at point assigned to each segmentation, with Kohli teaching the model at points with class labels and L the per class handling. Applicant’s argument regarding Ravanbakhsh is moot because Ravanbakhsh is no longer relied on. The rejection now relies on Azinovic. Azinovic and Kohli teaches coordinate used to obtain a class identifier. Regarding Applicant’s claim 22, Gormley teaches a beamforming system teaches a focused beam created by adjusting phase and amplitude at each antenna element with Desclos (US11742965B2) teaching unchanging antenna positioning. With Gormely and additionally Desclos, the amended claim 22 and new claim 24 is rejected. Lastly, every applied reference addresses the same recognizes problem, making a simulation of a physical scene grounded in real-world observations wherein the scene’s physical properties are partially known. For the above reasons, it is believed that the rejections should be sustained. Conclusion Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to DUNE NGUYEN whose telephone number is (571)272-8919. The examiner can normally be reached M-TH 7:00AM - 5:00PM. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Devona E Faulk can be reached at (571) 272-7515. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /DUNE NGOC NGUYEN/Examiner, Art Unit 2618 /DEVONA E FAULK/Supervisory Patent Examiner, Art Unit 2618