SYSTEM AND METHOD FOR RADIO AND OTHER PARAMETER ESTIMATION BASED ON MAPS

BACKGROUND:

Field:

Various communication systems may benefit from radio and other parameter estimations based on maps. For example, machine-type communication in long term evolution (LTE) communication systems may benefit from such estimations. The method presented here, according to certain embodiments, can be used outside of the area of radio for estimation of any location related parameters.

Description of the Related Art:

Generally, communication network operators may carry out many types of measurements concerning their networks. Many such measurements are included in the 3GPP standard specification, such as, for example, RSRQ (Reference Signal Received Quality) and RSSI (Received Signal Strength Indicator). These measurements may be based on a variety of factors including maps, parameter estimations, machine learning, communications, radio and navigation. Measurements concerning networks are often needed to improve the end user experience via more optimal network performance.

Obtaining network measurements may be assisted with the implementation of supervised, semi-supervised and unsupervised machine learning, which can be utilized to learn complex phenomena without explicitly having to model them. Further, nonlinear methods, such as deep learning are more powerful in terms of complexity of the modelled phenomena than linear methods, such as linear regression.

As wireless communication technology continues to develop, the amount of wireless data may increase 1,000 fold within the next 10 years. Essential elements in solving this challenge may include getting more spectrum, having smaller cell sizes and using improved technologies enabling more bits/s/Hz. With the amount of wireless data on the rise, it may also be expected that there will be at least an order of magnitude more connected devices.

The operation of networks will also get more complicated, and more measurement information will be needed due to the increase in the number of devices, more alternative frequencies to use and finer spatial resolution needed for measurements due to reduced cell sizes. As such, it may be beneficial to avoid carrying out at least some of the measurements, in order to save energy and other resources for communication.

SUMMARY:

According to certain embodiments, a method can include selecting at least one parameter dependent on location. The method can also include estimating values of the parameter at locations and times when measurements have not been carried out. The estimating can be based on map data, measured values of the parameter and other location dependent data combining different maps with overlapping coverage.

An apparatus according to certain embodiments can include at least one processor and at least one memory including computer program code. The at least one memory and the computer program code can be configured to, with the at least one processor, cause the apparatus at least to select at least one parameter dependent on location. The at least one memory and the computer program code can also be configured to, with the at least one processor, cause the apparatus at least to estimate values of the parameter at locations and times when measurements have not been carried out. The estimating can be based on map data, measured values of the parameter and other location dependent data combining different maps with overlapping coverage.

A computer program according to certain embodiments, can be embodied on a non- transitory computer readable medium, the computer program, when executed by a processor, can cause the processor to select at least one parameter dependent on location. The computer program, when executed by a processor, can also cause the processor to estimate values of the parameter at locations and times when measurements have not been carried out. The estimating can be based on map data, measured values of the parameter and other location dependent data combining different maps with overlapping coverage.

An apparatus according to other embodiments can include means for selecting at least one parameter dependent on location. The apparatus can also include means for estimating values of the parameter at locations and times when measurements have not been carried out. The estimating can be based on map data, measured values of the parameter and other location dependent data combining different maps with overlapping coverage.

BRIEF DESCRIPTION OF THE DRAWINGS:

For proper understanding of the invention, reference should be made to the accompanying drawings, wherein:

Figure 1 illustrates a flow diagram of obtaining enhanced estimates on label data in areas of interest, according to certain embodiments.

Figures 2A and 2B illustrate performance maps within a particular radio spectrum, according to certain embodiments.

Figure 3 illustrates a method according to certain embodiments.

Figure 4 illustrates another method according to certain embodiments.

Figure 5 illustrates standardization of inputs of an input area X, according to certain embodiments.

Figure 6 illustrates a prediction architecture, according to certain embodiments.

Figure 7 illustrates another method according to certain embodiments.

Figure 8 illustrates a system according to certain embodiments.

Figure 9 illustrates another method according to certain embodiments.

DETAILED DESCRIPTION:

Certain embodiments of the present invention may be applicable in performing radio and other parameter estimations based on maps. According to certain embodiments, a combination of measurement information existing as map-related label data, with different kinds of map-based data on one or more of satellite images, streets and buildings, terrain shapes, light detection and ranging (LIDA ) maps, installed network infrastructure and other available information can be performed. The combination of map-related label data and map-based data may be used in machine learning based methods or other estimation methods to estimate measurement values from locations and times when measurements have not been carried out. This general principle is illustrated in Figure 1.

Figure 1 illustrates a flow diagram 100 of obtaining enhanced estimates on label data in areas of interest according to certain embodiments. Input may include map-based data (maps) 105, represented by Xtrak, of an area, and map-related label data (for example, measurements) 110, represented by Ytrak, related to data such as radio performance measurements of an area. The map-based data 105 and map-related label data 110 may be fed into a training algorithm 115 for processing. The training algorithm 115 may include the implementation of deep learning, such as convolutional neural networks (CNN). Upon feeding the map-based data 105 and map-related label data 110 into the training algorithm 115, a trained model 120 can be established. The trained model 120 may be fed into a prediction algorithm 125, which may also include the implementation of deep learning, such as convolutional neural network. The trained model 120 may be described with W, which corresponds to all the parameters in all of the layers of the model. Along with the trained model 120, predicted map-based data 130 of an area(s), represented by X _pre dict, may also be fed to the prediction algorithm 125. The prediction algorithm 125 may incorporate the trained model 120 and the predicted map-based data 130 to generate enhanced estimates, such as map-related label estimates 135, represented by Y _pre dict, on the label data in the area(s) of interest. Such Y _pre dict values may correspond to radio performance estimates of an area.

Based on the general principle illustrated in Figure 1, it may be possible to obtain indications on the uncertainly of the estimates as a byproduct of the estimation method and, thus, provide a guide as to where and when to perform further measurements. The resulted estimates may be used for optimizing network performance. Further, the resulted estimates may also be used for optimizing connectivity or routes.

According to certain embodiments, different kinds of maps may be combined with parameter estimation values to create estimates on other measurements. This may include, for example, combining maps and information not visible in the maps in order to produce better estimates. The indications of the uncertainly of the estimates from the combination of the different kinds of maps with parameter estimation values may be used to guide further measurements. For example, in some area(s) where the obtained estimates have high uncertainly, it would be useful to carry out real measurements of those areas. Some estimates could also show very bad results for end users. In such a case, it could be beneficial to carry out more measurements to confirm the situation and to be able to search for as much proper improvement as possible.

According to other embodiments, the combination of the different maps with parameter estimation values may be used to create performance maps for different portions of a spectrum of a wideband radio system where conditions may be quite different for different portions of the spectrum. For example, a wideband radio system may cover a frequency range of 1 to 90 GHz.

According to other embodiments, estimation of location related parameters in fields outside of radio may be possible. For example, maps with information on infrastructure, such as roads, and information on societal factors, such as education, age, type of cars registered to addresses in an area can be used to predict probabilities for various types of traffic accidents.

In certain embodiments, the parameter estimation values may relate to performance metrics (map-related label data). The performance metrics may include data gathered from reference signal received power ( SRP) or RSRQ, signal-to-interference-plus-noise ratio (SINR), spectral efficiency, or throughput. Further, the maps (map-based data) may include information regarding terrain, surroundings, buildings, weather, foliage and load metric. The surroundings may include, for example, trees, map posts, etc. Related information to the maps may include geographical and terrain information as "visible" to usage of radio, including access point location, obstacles such as trees, pillars and other objects that may influence propagation and line of sight conditions. Related information to the maps may also include user information such as, for example, user location, movement, tracks and predicted tracks of movement. Related information to the maps may further include weather and seasonal information as "visible" to usage of radio, including foliage, rain or snow.

These maps may be somewhat semi-static. For example, the maps may be slowly changing. According to certain embodiments, the maps may be changing in order of hours. The maps may also be generated, refined and updated via periodic feedback. Γη certain embodiments, the periodic feedback may be derived from user equipment (UEs). For example, the feedback may be based on channel quality indication (CQI) or RSRP/RSRQ. The network may take semi-static maps and combine them with dynamic information such as available bandwidth from sensing and user specific information including, for example, location, speed, and direction.

According to certain embodiments, the map information, or at least part of the information, acquired by the combination of different kinds of maps with parameter estimation values may be provided in association with other maps and services. The information may also be valuable for local and small operators and for novel frequency bands.

An example of semi-static performance maps is shown in Figures 2A and 2B for two frequency bands, one for 885 MHz (Figure 2A), and one for 2.6 GHz (Figure 2B). As shown in Figures 2A and 2B, downlink SINR from the best serving cell is shown and used as the performance metric. From the maps, it can be seen that performance can be quite different at different bands. According to certain embodiments, the network can estimate such maps based on partial information. Γη other embodiments, the combination of different maps with radio parameter estimation values may be used for estimating other parameters than those related to radio. Deep learning may have good capabilities to combine multimodal information such as case maps and radio data.

In certain embodiments, information on the maps may need to be dependent on the technology to be used for wireless connectivity, if such a dependency exists for the particular information. Further, a central entity in the network such as, for example, a macro base station, may be carrying out the estimations. This may also be done higher in the hierarchy. According to certain embodiments, the estimations could be combined with map information (outdoor and/or indoor) including the type of the terrain, and potential blocking elements for line-of-sight (LOS) communication. Such information may include trees, pillars, other supporting elements and other objects with reasonable heights.

The impact of current weather and climate on coverage may also be incorporated into the map information by updating it at appropriate periods with feedback from UEs. Such weather conditions may correspond to changes with the time of year and type of trees, and may include fog, rain, snow and foliage.

According to certain embodiments, the network may select different portions of the spectrum for the purpose of creating performance maps. This selection can be based on radio propagation characteristics, pathloss exponents, line-of-sight requirements, oxygen and water absorption characteristics, and penetration characteristics. In other embodiments, the network or UE may transmit signals such as pilots, beacons, or synchronization signals for measurement in different portions of the spectrum. Alternately, wideband signals covering the entire spectrum of interest can be used. According to other embodiments, the network may use measurements in one portion of the spectrum to estimate the performance in another portion of the spectrum. This can be done, for example, through adjusting and applying different pathloss exponents to measurements from one portion of the spectrum. Pathloss exponents describe the strength of the decay of the signal over distance. Such adjustment can be based on known measurements on the change of the parameter values as a function of each other. For example, there may be measurements on absorption per distance as a function of frequency. In such graphs, there may be a peak of absorption around 60 GHz due to the physical characteristics of oxygen. Due to this knowledge, the selection of the spectrum portion can be based on including or excluding this peak area due to similarities with the portion of the spectrum under interest. Due to such differences, there can be requirements on line-of-sight conditions for the connectivity, depending on the technology and the physical characteristics of the radio propagation at different frequencies. Therefore the selection of the portion of spectrum can also be dependent on the need or lack of need for line-of-sight in the different portions of the spectrum. Thus, according to certain embodiments, in general, the selection of the portion of spectrum can be strongly related to similarities or differences in the behaviour the parameters of interest as a function of frequency.

According to other embodiments, the network may adjust the performance maps based on expected environmental changes such as weather conditions (for example., rain or snow) and vehicular and pedestrian traffic density (for example, rush hour, train arrival, etc.). Further, in other embodiments, the network may adjust the performance maps based on spectrum usage by the network. For example, as spectrum is dynamically allocated by the network, the maps may be adjusted to reflect this usage including bandwidth availability, interference, and load.

In certain embodiments, the tasks described above can be mapped to unsupervised, semi-supervised, or supervised machine learning tasks. In a supervised machine learning setting, the input X can be, for example, the satellite image, LIDA point cloud, a map in a raster image representation or a combination thereof. Generally, the input X corresponds to input data without labels. The input X may also correspond to both Xpredict and Xtrain, depending on whether the model is being trained, or whether predictions are being made with the model. For example, Xpredict may be used when predicting, and Xtrain may be used when training. The trained and predicted labels Y may correspond to the image representations of the target quantities of interest such as, for example, network performance in that location of the map. In a semi- supervised setting, the network may first be pre-trained using X and then fine-tuned using X and Y. The semi-supervised setting may allow for use of larger datasets in case labels Y are scarce. In an unsupervised setting, labels may not be used at all, but the different parts of the map X may be clustered based on some similarity measurements. Further, depending on the predicted quantity, the system can be run in a classification or regression mode.

According to certain embodiments, the data X and Y may be split in training instances so that the input Xi for each training instance may be a rectangular area of k _x i*k _X 2 pixels. The corresponding label Yi may be a rectangular area of k _y i * k _y2 pixels, corresponding to the center of Xi. Further, k _y <= k _x , and typically much smaller than

In certain embodiments, deep neural networks, for instance deep convolutional neural networks can be utilized in the tasks and machine learning based methods or other estimation methods described above. Deep neural networks perform well in image and map-related tasks. In deep neural networks, multiple layers of neural networks with weights W may be combined one after another. Each layer may also apply a non-linear function such as a hyperbolic tangent (tanh) or a rectified linear unit ( eLU). In convolutional neural networks, some of the layers may be convolutional (for example, applying a set of 2D convolutional filters to the input), and some may be pooling layers that subsample the input. In other embodiments, various other techniques may be implemented. Some of these techniques are described in Jarrett, Kevin, et al, "What is the best multi-stage architecture for object recognition?" Computer Vision, 2009 IEEE 12th International Conference on. IEEE, 2009, the contents of which are hereby incorporated herein in their entirety. Other techniques are described in LeCun, Yann, et al, "Convolutional networks and applications in vision," Circuits and Systems (ISCAS), Proceedings of 2010 IEEE International Symposium on. IEEE, 2010, the contents of which are hereby incorporated herein in their entirety. Techniques are also described in Y. LeCun, et al, "Handwritten digit recognition with a back-propagation network," in NIPS '89, the contents of which are hereby incorporated herein in their entirety. Figure 3 illustrates a method of training the deep neural networks, according to certain embodiments. Other methods may be found in Y. LeCun, discussed above. In particular, the weights W may be optimized so that certain loss function is minimized. This may be accomplished by so-called backpropagation algorithm, which works by computing a gradient of the loss function, with respect to all parameters in W, as shown at 301. Further, at 305, the parameters are adjusted accordingly from a forward pass computation from the Xtrain. In classification, an example of a loss function may be the cross-entropy loss as used in multiclass logistic regression. Different optimization methods can be utilized to train the networks. One example is batch stochastic gradient descent (SGD). In stochastic gradient descent, the various patches Xi, Yi may be represented randomly in small patches and the weights W may slowly change towards a good local or global minimum.

Figure 4 illustrates a method of predicting the value of interest Y _pre dict for X _pre dict, according to certain embodiments. In particular, this may be accomplished by presenting the input X _pre dict to the network, as shown at 401. At 405, the value Ypredict can be computed using the trained values of W, and at 410, a forward pass computation can be performed to compute Ypredict. For instance, in a simplified two layer network with one convolutional and one fully connected layer of certain embodiments, Y = W& ^■ relu((W ^k * X) + bk) +bf _C can be computed, where * is the 2D discrete convolution operator, bk and b& are trainable bias parameters, W ^k are the trainable convolution filters of given size (si,s ₂ ), and ^■ is the dot product of two matrices. All parameters are matrices and their sizes may depend on size of input, output and the layers. The number of filters k may determine the dimensionality of the output of that layer. According to certain embodiments, k may correspond to channels. The function relu(x) corresponds to the function y=max(0,x) applied elementwise. In real world use cases, many layers may be used instead of two.

Deep learning results can be improved by combining, for example, dropout for regularization and adaptive learning rate algorithm (AdaDelta) or momentum method for adaptive learning rate with the method. An example of AdaDelta is described in Zeiler, Matthew D., "ADADELTA: An adaptive learning rate method," arXiv preprint aXiv: 1212.5701 (2012), the contents of which are hereby incorporated herein in their entirety. There are also various other existing methods that can be used in combination.

For time-dependent data, recurrent neural networks (RNN) or a combination of recurrent neural networks and convolutional neural networks can be utilized.

As shown in Figure 5, since the propagation of radio signals depend partly on the geographical elements between the sender and the receiver, the results can be improved by standardizing input area X so that both the sender and the receiver always lie within the area X. With reference to Figure 5, S refers to a sender, Rl and R2 refer to receives, w refers to width, and h refers to the height/length of the area.

According to certain embodiments, the input patch can be standardized into a rectangle where the sender and receiver are always in the same pixel positions, but the resolution changes based on the distance between the sender and the receiver. The distance between the sender and the receiver can be inputted as separate variables to the process. Additionally, the patches can be rotated to be of standard orientation.

According to other embodiments, the input patch may be standardized differently so that the receiver is always in the same position and the sender's position can change, but always within the input (the input size is selected so that the sender is always visible when performing predictions at all locations of interest). Further, the resolution may always be kept the same. The orientation can be either kept or can be rotated. If the orientation is kept, the receiver would be placed in the center of the patch so that the prediction output would match it.

In other embodiments, the sender may be considered to be a grid, and the estimation may not take a specific sender position into account. According to certain embodiments, the different kinds of maps and parameter estimation values, such as, for example, values related to radio, can be combined in a variety of ways, such as the following manner. In particular, input data X can be a three-dimensional array (or tensor), where the first two dimensions represent the pixel locations. Each matrix in the third dimension can be unrelated, but pertain to location-specific data. For example, according to certain embodiments, the first three matrices could be the red, green blue (RGB) values of a satellite image for each pixel. The next matrices could be, for example, building and road structure heights that are derived for each pixel from a map database. The next matrices could be the elevation data that have been acquired, for example, using aerial lidar. There can also be radio- related measurements in X. For instance, existing measurements of background noise in certain frequency channels that are inputted, for example, by interpolation for each pixel in X. In deep learning cases, each neuron (or convolutional filter) in the second matrix may be connected to all of the matrices in the input X, although it may be beneficial to connect different types of input only at higher levels of the network. Y can be any radio related property that has been measured in some areas and need to be predicted in other areas.

Figure 6 illustrates a prediction architecture when multiple types of map-related input data are used, and all input data is normalized spatially so that a value may be given for each pixel in the map. In particular, Figure 6 illustrates a concrete example of Figure 1 and Figure 5 in more detail. The input X 605 may be a three-dimensional array (or tensor) where the first two dimensions are the 2D coordinates and the matrices in the third dimension represent the different input data modalities. The different data input modalities may include, for example, satellite images, elevation maps and background noise measurements. Each of the matrices may have the same normalized width and height (Wstd, Hstd). The sender (S) and the receiver (R) are marked in Figure 4 with dots. The input X 605 may be fed into a deep neural network 610, which consists of multiple neural layers 615, 620, 625, 630 and 635, corresponding to the training and prediction algorithms 115, 125 in Figure 1. The convolution layers 615, 625, and 630 compute the function yk =relu((W ^k * x) + bk), where * is the 2D discrete convolution operator and bk is a trainable bias parameter, and W ^k are the trainable convolution filters of given size (si,s ₂ ). The number of filters (i.e., channels) determines the dimensionality of the output of that layer. The function relu(x) corresponds to the function y=max(0,x) applied element wise. The subsampling layer 620 reduces the dimensionality without any trainable parameters, and its output is given by the maximum activation over non-overlapping rectangular regions of given size (si, s ₂ ). The fully connected layer 635 is normal neural network layer thus, computing the function Y = relu(Wx +b). The output Y is the measure that is to be predicted, in this example, the signal strength at the receiver ( ).

At training time, both input Xtrain and output Ytrain (see Figure 1) may be fixed as the free parameters of the network are estimated using the training algorithm 115. The estimated parameters may be called the "Trained Model" in Figure 1. During training, all different (Xtrain, Ytrain) pairs may be presented to the training algorithm 115. After training, the prediction algorithm 125 may utilize the Trained Model to compute estimated output value Ypredkt for signal strength at any location, as corresponding area X _pre dict is presented as input.

According to certain embodiments, the estimation of the real value output Y may be performed with regression. According to other embodiments, the value to be estimated may be binned to N bins and the problem may be to set up as a classification task, which allows output of the probability distribution of the value within the bins and, thus, compute the estimates of the uncertainly of the estimated output value. For example, in certain embodiments, using logistic regression allows for the computation of uncertainly estimates in this fashion.

Examples of other supervised or semi-supervised machine learning methods for these tasks include, but are not limited to, random forest, support vector machine (SVM), Naive Bays and Gaussian processes. Examples of other unsupervised machine learning methods applicable for these tasks may include, but are not limited to, K-Means clustering, deep auto encoders and restricted bolzmann machines. Further, training and prediction phases can be run in a single or multiple central processing unit (CPU) or graphics processing unit (GPU).

Figure 7 illustrates a method according to certain embodiments. Figure 7 more specifically illustrates a process of creating a performance map and measurement estimates. As shown in Figure 7, the method may include, at 701, selecting at least one parameter of interest having dependency on location. The method may also include, at 705, creating a performance map for the at least one parameter of interest by combining a plurality of different maps of at least overlapping coverage area with at least one parameter estimation value for at least one location. The method may further include, at 710, creating estimates on other measurement values from locations and times when measurements have not been carried out. The estimates on the measurements may be based on the combination of the different maps with the parameter estimation values.

Figure 8 illustrates a system according to certain embodiments. In one embodiment, a system may include multiple devices, such as, for example, at least one UE 810 and at least one eNB 820, other base stations or access points, or any device with transmitter or receiver functionality. In certain systems, only UE 810 and eNB 820 may be present, and in other systems, UE 810, eNB 820, and a plurality of other user equipment may be present. Other configurations are also possible.

Each of these devices may include at least one processor, respectively indicated as 814 and 824. At least one memory can be provided in each device, as indicated at 815 and 825, respectively. The memory may include computer program instructions or computer code contained therein. The processors 814 and 824, and memories 815 and 825, or a subset thereof, can be configured to provide means corresponding to the various blocks of Figures 1, 3, 4, 6 and 7, and any other method(s) and computational step(s) described above. Although not shown, the devices may also include positioning hardware, such as a global positioning system (GPS) or micro electrical mechanical system (MEMS) hardware, which can be used to determine location of the device. Other sensors are also permitted and can be included to determine location, elevation, orientation, and so forth, such as barometers, compasses, and the like.

As shown in Figure 8, transceivers 816 and 826 can be provided, and each device may also include at least one antenna, respectively illustrated as 817 and 827. The device may have many antennas, such as an array of antennas configured for multiple input multiple output (MIMO) communications, or multiple antennas for multiple radio access technologies. Other configurations of these devices, for example, may be provided.

Transceivers 816 and 826 can each, independently, be a transmitter, a receiver, or both a transmitter and a receiver, or a unit or device that is configured both for transmission and reception.

Processors 814 and 824 can be embodied by any computational or data processing device, such as a central processing unit (CPU), application specific integrated circuit (ASIC), or comparable device. The processors can be implemented as a single controller, or a plurality of controllers or processors.

Memories 815 and 825 can independently be any suitable storage device, such as a non- transitory computer-readable medium. A hard disk drive (HDD), random access memory (RAM), flash memory, or other suitable memory can be used. The memories can be combined on a single integrated circuit as the processor, or may be separate from the one or more processors. Furthermore, the computer program instructions stored in the memory and which may be processed by the processors can be any suitable form of computer program code, for example, a compiled or interpreted computer program written in any suitable programming language.

The memory and the computer program instructions can be configured, with the processor for the particular device, to cause a hardware apparatus such as UE 810 and eNB 820, to perform any of the processes described above (see, for example, Figures 1, 3, 4, 6 and 7). Therefore, in certain embodiments, a non-transitory computer-readable medium can be encoded with computer instructions that, when executed in hardware, perform a process such as one of the processes described herein. Alternatively, certain embodiments of the invention can be performed entirely in hardware.

Furthermore, although Figure 8 illustrates a system including a UE 810 and an eNB 820, embodiments of the present invention may be applicable to other configurations, and configurations involving additional elements.

Figure 9 illustrates a method according to certain embodiments. Figure 9 more specifically illustrates a method for creating a performance map. As shown in Figure 9, the method may include, at 901, selecting at least one parameter dependent on location. According to certain embodiments, the selection may include selecting a portion of a spectrum in a wideband radio system 955. According to other embodiments, the parameter may correspond to various measurable data. For example, the parameter may correspond to a general performance metric, Y, which may be dependent on, or related to, other location dependent information, X.

The method may also include, at 905, estimating predicted values of the parameter at locations and times when measurements have not been carried out. The estimating can be based on map data, measured values of the parameter and other location dependent data, Xpredkt. According to certain embodiments, the predicted values of the parameter may correspond to map-related label estimates. For example, the predicted values may correspond to the map-related label estimates 135, Ypredkt, as shown in Figure 1. Further, the locations and times when measurements have not been measured may correspond to map-based data. For example, the locations and times when measurements have not been measured may correspond to the map-based data 130, Xpredict, as shown in Figure 1. Additionally, the map data may correspond to map-based data 105, Xtrain, and the measured values of the parameter may correspond to map-related label data 1 10, Ytrain, as shown in Figure 1. The method may also include, at 910, creating a performance map for the parameter by combining the measured values of the parameter, the map data, and the predicted values of the parameter. According to certain embodiments, the performance map may generally refer to any location based map. For example, the performance map may refer to any location based radio map.

The method may further include, at 915, using the estimated values to guide further measurements of measured values. The method may also include, at 920, obtaining dynamic information from sensing and user specific information. The method may further include, at 925, combining the dynamic information with at least one of the plurality of different maps.

The method may also include, at 930, adjusting the performance map. According to certain embodiments, the performance map may be adjusted in various ways. For example, the performance map may be adjusted based on expected environmental changes 935. In other embodiments, the performance map may be adjusted based on vehicular and pedestrian traffic density 940. In other embodiments, the performance map may be adjusted based on spectrum usage 945. The method may further include, at 950, using measurements in one portion of the spectrum to estimate performance in another portion of the spectrum.

According to certain embodiments, various advantages may be achieved. For example, better end user experience and enhanced network operation via better information for the network to optimize its operation, and optimal use of network resources may be achieved. Further, optimal performance based on the circumstances, and minimization and optimization of real measurements needed, may result in energy and cost savings. Certain embodiments may provide information for measurements in areas where measurements normally cannot be made. Further, given enough data, deep learning can learn to model complex phenomena, and it can also use raw untransformed data, such as pixel intensities as input. Other embodiments may provide facilitated network planning. One having ordinary skill in the art will readily understand that the invention as discussed above may be practiced with steps in a different order, and/or with hardware elements in configurations which are different than those which are disclosed. Therefore, although the invention has been described based upon these preferred embodiments, it would be apparent to those of skill in the art that certain modifications, variations, and alternative constructions would be apparent, while remaining within the spirit and scope of the invention. In order to determine the metes and bounds of the invention, therefore, reference should be made to the appended claims.

Glossary

ASIC Application Specific Integrated Circuit

CNN Convolutional Neural Networks

CPU Central Processing Unit

CQI Channel Quality Indication

GPS Global Positioning System

GPU Graphics Processing Unit

HDD Hard Disk Drive

LTE Long Term Evolution

LIDA Light Detection and Ranging

LOS Line-of-sight

MEMS Micro Electrical Mechanical System

MIMO Multiple Input Multiple Output

RAM Random Access Memory

ReLU Rectified Linear Unit

RGB Red Green Blue

RNN Recurrent Neural Networks

RSRP Reference Signal Received Power

RSRQ Reference Signal Received Quality

RSSI Received Signal Strength Indicator SINR Signal-to-interference-plus-noise Ratio

SVM Support Vector Machine