Forecasting radio spectrum use

Abstract

An example embodiment describes a computer-implemented method for predicting radio spectrum usage in an environment including interference (110-111, 210-211) and a plurality of wireless nodes (100-105, 200-206) by a neural network, the nodes (100-105, 200-206) operating within the radio spectrum via a slotted medium access control (MAC) protocol, the method comprising the steps of obtaining radio spectrum usage for each time slot by sensing the radio spectrum; marking a time slot as usable when the corresponding radio spectrum usage is below a predefined noise threshold or when the node has successfully transmitted a message through it, and marking a time slot as unusable when the slot is not usable and suitable for receiving a message; and calculating a loss function as a cross entropy loss of usable time slots and unusable time slots divided by the total number of usable time slots and unusable time slots, and training the neural network by minimizing the loss function.

Description

Technical Field

[0001] The present invention generally relates to the field of predicting radio spectrum usage in an environment including interference. Background Art

[0002] Wireless spectrum is becoming a critical resource for many of today's devices. More and more devices are connecting to the internet via wireless radios and transmitting increasing amounts of data using many different wireless networking technologies. Most of these technologies still use the Industrial, Scientific, and Medical (ISM) radio spectrum. The scarcity of radio frequency bands and spectrum is one of the most significant challenges now and in the future.

[0003] Today’s existing collision avoidance techniques either apply random backoffs when spectrum collisions are detected or assume knowledge of other nodes’ spectrum occupancy. These approaches are excellent solutions for optimizing spectrum usage between networks, but fail when it comes to optimizing the overall spectrum usage and throughput of all neighboring wireless networks.

[0004] Since most existing algorithms optimize their own network performance, and since the number of different technologies using the same ISM band is growing, the algorithms compete with each other, resulting in an overall degradation of all networks. In addition to other wireless networks, the environment may also contain interference (such as microwaves) that can affect the overall performance, although these interferences are uncontrollable.

[0005] It is therefore an object of the present invention to alleviate the above mentioned drawbacks and to provide an improved solution to predict spectrum usage in an environment comprising one or more wireless networks and one or more interferences. Summary of the Invention

[0006] In a first aspect, the object is achieved by a computer-implemented method for predicting radio spectrum usage in an environment comprising interference and a plurality of wireless nodes operating via a slotted medium access control (MAC) protocol within the radio spectrum by means of a neural network, the method comprising the following steps:

[0007] - Obtain radio spectrum usage for each time slot by sensing the radio spectrum;

[0008] -Mark time slots:

[0009] Mark a slot as available when the corresponding radio spectrum usage falls below a predefined noise threshold or when a node has successfully transmitted a message through it;

[0010] o Marking a timeslot as unusable when it is not available and suitable for receiving a message; and

[0011] - calculating the loss function as the cross entropy loss of usable and unusable time slots divided by the total number of usable and unusable time slots; and

[0012] - Train the neural network by minimizing the loss function.

[0013] The environment includes multiple wireless nodes configured to communicate with each other, thereby forming a wireless network. Interference also exists within the environment. For example, such interference may be other wireless nodes in another wireless network. Interference may also be devices that transmit radio signals within a frequency range that interferes with the frequency range of the wireless network. An example of such a device is a microwave oven. The group of unknown networks and / or other sources may also be designated as an interfering network cluster INC.

[0014] Within the environment, there are therefore multiple technologies with a wide variety of different applications. Furthermore, nodes can also be mobile, which makes the environment very dynamic.

[0015] Wireless nodes and interferers share the radio spectrum, or at least have overlapping radio spectrum ranges used by the wireless nodes. These wireless nodes use the radio spectrum to operate, i.e., exchange data and / or data packets. To exchange data packets, the wireless nodes are configured to operate via a slotted MAC protocol. This is accomplished using a scheduler technique, which can be centralized or decentralized.

[0016] The time-slotted MAC protocol is, for example, a time division multiple access (TDMA) protocol, a multi-frequency time division multiple access (MF-TDMA) protocol, or a time-slotted carrier sense multiple access (CSMA) protocol.

[0017] A node has access to one or more channels operating within the radio spectrum. Time slots are used to exchange data packets across the channels. Time slots can be grouped into superframes or within created time slot windows. It should also be understood that the grouping of time slots depends on the protocol used. Because nodes operate within their own network using the slotted MAC protocol, each node can perform actions simultaneously with other nodes in the network. For example, an action could be receiving or transmitting a data packet.

[0018] The reception and transmission of data packets are performed using time slots. Since the network includes INC, a time slot may be occupied by one or more devices in INC and thus unavailable to the node. In addition, the technology used by the devices in INC is unknown.

[0019] Therefore, as a first step, the radio spectrum usage for each time slot is determined by sensing the radio spectrum. This sensing can be performed, for example, by hardware radio technology or by a software-defined radio (SDR) platform. Sensing measures the energy in the radio spectrum. The energy of each channel in the radio spectrum is sensed, thereby determining the energy level for each time slot.

[0020] Next, since nodes themselves also use time slots to transmit messages, the sensed energy may originate from one or more of these nodes. To distinguish between energy sensed from the INC and from one or more nodes, nodes are configured to identify whether the current time slot is allocated within the network. In the MF-TDMA protocol, this information is known to the nodes, and for the CSMA protocol, this can be achieved by using a Request-to-Send / Clear-to-Send (RTS / CTS) strategy.

[0021] Next, since it's possible to distinguish between the INC's sensed energy and energy originating from the node itself, a timeslot can be marked as usable when the corresponding radio spectrum usage falls below a predefined threshold or when the slot is used by one of the nodes to successfully transmit a message. In other words, when the energy level is low, the timeslot can be used by the network because the INC will not interfere with it. Secondly, it remains usable when it has already been successfully used for transmission.

[0022] When other time slots are not marked as usable but are suitable for receiving messages, they may be marked as unusable.

[0023] In case they are remaining time slots that are not marked as usable or unusable, these remaining time slots may be marked as unknown, or even not marked at all and further ignored.

[0024] Then, using the usable time slots and the unusable time slots, a loss function is calculated as the cross entropy loss divided by the total number of usable time slots and unusable time slots.

[0025] Finally, a neural network configured to predict radio spectrum usage in the environment is trained by minimizing a loss function.

[0026] Different advantages can be identified. First, by predicting spectrum usage at each node for the upcoming time slot, learning is possible in unknown environments. Second, through the loss function, predictions based on partially observable data can be optimized. This makes the predictions robust to unknown behaviors and novel scenarios. Third, through predictions, conflicts with other sources or interference can be avoided.

[0027] According to an embodiment, the computer-implemented method further comprises the step of optimizing the layers of the neural network.

[0028] For example, the optimization may be performed by backpropagation, Tabu search, and / or simulated annealing.

[0029] Optimization can also be performed by the swish optimization function. The swish activation function is a smooth non-monotonic function defined as f(x) = x.σ(x), where σ(x) = (1 + exp(-x)) -1 The Swish activation function improves the accuracy compared to the traditional ReLU activation function.

[0030] According to an embodiment, the optimizing further comprises optimizing an output layer of the neural network by means of a softmax activation.

[0031] In other words, the swish activation function is used on every layer of the neural network except the last output layer. For the output layer, a softmax activation is applied. This normalizes the output into a probability distribution that includes multiple probabilities proportional to the exponent of the input number.

[0032] According to an embodiment, the computer-implemented method further comprises the step of predicting interfering upcoming radio spectrum usage by means of a neural network.

[0033] As the neural network is trained, it becomes suitable for predicting upcoming radio spectrum usage by interferers (and therefore INC’s devices).

[0034] According to a second aspect, the present disclosure relates to a neural network according to the first aspect.

[0035] For example, a neural network may be trained in an environment or in an environment that is configured to simulate a representative environment in which the neural network will be used.

[0036] According to a third aspect, the present disclosure relates to a data processing system comprising means for performing the method according to the first aspect.

[0037] According to a fourth aspect, the present disclosure relates to a computer program product comprising computer executable instructions for performing the method according to the first aspect when the program is run on a computer.

[0038] According to a fifth aspect, the present disclosure relates to a computer-readable storage medium comprising the computer program product according to the fourth aspect.

[0039] According to a sixth aspect, the present disclosure relates to a node for operating in a wireless network, the node being configured to predict radio spectrum usage in an environment by means of a computer-implemented method according to the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Some example embodiments will now be described with reference to the accompanying drawings.

[0041] Figure 1 An example embodiment of a ring topology of a network with known and unknown networks is shown;

[0042] Figure 2 Another example embodiment of an access point AP topology for a network having known networks and unknown networks is shown;

[0043] Figure 3 The average total throughput of superframes in multi-frequency time division multiple access MF-TDMA, mesh network and interference network cluster INC for different scheduling methods is shown;

[0044] Figure 4 The network metrics of superframes in a MF-TDMA mesh network for different scheduling methods are shown;

[0045] Figure 5 The average total throughput of superframes in MF-TDMA AP network and INC for different scheduling methods is shown;

[0046] Figure 6 The network metrics of superframes in a MF-TDMA AP network for different scheduling methods are shown;

[0047] Figure 7 The average total throughput of superframes in carrier sense multiple access (CSMA), AP network, and INC for different scheduling methods is shown;

[0048] Figure 8 shows the network metrics of superframes in a CSMA AP network for different scheduling methods;

[0049] Figure 9 shows an example embodiment of a suitable computing system 900 for performing one or more steps in embodiments of the present invention; and

[0050] Figure 10 An example embodiment of the architecture of the algorithm of the invention is shown for a self-network comprising three nodes. DETAILED DESCRIPTION

[0051] exist Figure 1 In FIG, a ring topology of a network is illustrated. The topology includes a known network including nodes 100-105 configured to communicate with each other. The topology also includes a second unknown network including nodes 110-11, which are also configured to communicate with each other but not with other nodes 100-105. Figure 2Another topology is illustrated in FIG. This topology includes nodes 200-206 arranged in an access point (AP) topology. Other nodes 210-211 illustrate a second unknown network. Nodes 110-111 and 210-211 illustrate an interfering network cluster INC. INC interferes with the networks of nodes 100-105 and 200-206. Thus, a known network includes a plurality of known nodes N, representing the own network, and at least a second group of unknown networks INC. INC may include noise from other sources (e.g., microwaves).

[0052] Nodes 100-105 and 200-206 of their own network can access the C channel and use the Slotted Medium Access Control (MAC) protocol. Examples of the MAC protocol include the Multi-Frequency Time Division Multiple Access (MF-TDMA) protocol or the Slotted Random Access (SRA) protocol. The superframe used in the Slotted MAC protocol consists of S time slots. Therefore, when using the MF-TDMA protocol, the superframe can be equal to the MF-TDMA superframe, or, if using the SRA, it can be an artificially created window of S time slots.

[0053] Furthermore, t is defined as a global reference of time, where t is the number of slots executed after the first radio in the environment is started. Since all nodes 100-105 and 200-206 use the slotted MAC protocol, all nodes 100-105 and 200-206 execute exactly one slot at the same time, as this is a general property of the slotted MAC protocol. Node n among nodes 100-105 and 200-206 executes action on each channel c at time t also, and is represented as:

[0054]

[0055]

[0056] There is no constraint on the number of actions a node can execute at a particular time t. As long as they are executed on different channels. This is expressed as:

[0057]

[0058] In addition, INC is completely unknown, so there may be an unknown number of nodes in INC, using unknown protocols, or even the noise generator may be part of INC. For INC, P t,c is defined as the usage of channel c by INC at time t, where if a source in INC uses channel c at time t, then P t,c =1, otherwise P t,c= 0. Due to factors such as fading, power control, modulation and coding scheme MCS, settings, etc., the interference of the generated signal may be different for different nodes at different locations. Therefore, is defined as the case where INC in the interpretation of node n uses channel c at time t with energy above a predefined noise threshold. Where, if a source from INC as the interpretation of node n produces energy above the noise threshold on channel c at time t, then

[0059] To make better and more intelligent use of the spectrum, the MAC protocol in the own network avoids collisions, and for each transmission, at least one of the nodes 100-105 and 200-206 listens. This can be achieved, for example, using a centralized MF-TDMA algorithm. Communication between different nodes in the own network is successful if, at a given time slot t, at least one node is in the RX state and receives messages from only one other node in the own network on a given channel c:

[0060]

[0061] Therefore, if a self-node transmits on channel c where the destination node is not affected by INC, then the communication of the self-network on channel c at time t is successful. t,c Used to determine whether its own message can be successfully delivered:

[0062]

[0063] Because INC is unknown, and since from an external perspective it is not possible to know whether a transmission interferes with INC messages, if γ′ t,c =1, then it can be assumed that the communication in INC is successful, where:

[0064]

[0065] In addition, the maximum network packet throughput is the sum of all successful transfers:

[0066]

[0067] Similarly, in this way, the maximum network packet throughput of INC is:

[0068]

[0069] In addition, Γ and Γ′ are defined as the actual packet throughput of the own network and INC, respectively.

[0070] By the present invention, the total network throughput of all networks is maximized, both the own network and the INC, by selecting the best possible action A for all channel-time pairs for all nodes n in the own network. The total network throughput of a superframe is defined as:

[0071] F(α)=αΓ+(1-α)Γ′ where α∈[0, 1] (Equation 9)

[0072] Where α represents the relative priority between the own network and INC.

[0073] To increase throughput Γ and Γ′, either the amount of transmission (TX / RX operations) is increased, more data is generated, or the number of collisions between the INC and its own network is reduced. Since there is no control over data generation and only the nodes 100-105 and 200-206 in its own network are controlled, throughput is increased by reducing collisions.

[0074] If there is a potential collision between INC and its own network on channel c at time slot t, then the parameter ω t,c is defined as 1, and Ω is the total number of potential conflicts:

[0075]

[0076]

[0077] According to Equation 4, it can be assumed that there is no collision in the own network, so in order to maximize F(α), Ω needs to be minimized. In addition, each node n can measure the energy on all c channels based on hardware technology (such as software defined radio SDR).

[0078] In addition, the parameter is defined as the observation of channel c by node n at time t. If the amount of energy detected on channel c by node n at time t is too high to receive another message, then the parameter If no energy is detected on the channel or a small amount of energy is detected and does not affect the successful reception of the message, then the parameter Detect all energy, including energy generated by the node itself or neighboring nodes in its own network.

[0079] Each node contains network information and statistics, i.e., the node knows whether the current time slot is allocated within its own network, where if the time slot on channel c at time t will be used by any node in its own network, then TX t,c, = 1, as described in Equation 12 below. For example, this can be achieved by a request to send / clear to send RTS / CTS strategy. In addition, each node knows whether it has received the message correctly. And when node n correctly receives the message at time t on channel c,

[0080]

[0081] like Figure 10 As shown in FIG, the method of the present invention includes five components: a spectrum monitor 1000, a pre-processing unit 1001, a predictor unit 1002, a probability matrix 1003 on a node, and a centralized or decentralized overall scheduler 1010. The components are present at each node, such as node A, node B 1005, and node C 1006. These components will now be discussed further with reference to node A.

[0082] The spectrum monitor 1000 captures the energy of the entire spectrum. The captured spectrum monitor 100 information is forwarded to the pre-processing unit 1001. The pre-processing unit 1001 of node A creates correct observations for node A and all t in the most recent superframe on each channel c. and network scheduling TX t,c , node scheduling and success message

[0083] For each superframe, the predictor unit 1002 predicts the upcoming spectrum usage that will be generated by the INC. Parameters is defined as the result of prediction unit 1002, where It describes whether the time slot will be predicted to be highly used (values close to zero) or will be predicted to be idle (values close to one). Parameter The values of the probability matrix are called probabilities. Scheduler 1010 uses probability matrix 1003 to select free time slots to avoid collisions. Scheduler 1010 can be a centralized or decentralized MF-TDMA scheduler, or a CSMA AP that defines the time slots when nodes can transmit. This can be implemented in an MF-TDMA network, where a sending node requests a time slot from a receiver, which responds with the best available time slot based on a prediction.

[0084] Thus, the predictor unit 1002 predicts the used time slots in the upcoming superframe to be used by the INC. These predictions can be used as actions for the scheduler 1010 to select time slots. The prediction is defined as As a strategy, the prediction unit 1002 of node A is based on the observation matrix O at time t n and the scheduling matrix TX prediction, where H defines the number of historical frames. In addition, π *n (O n ,TX) is defined as the theoretically optimal strategy.

[0085] This problem can also be modeled as a reinforcement learning (RL) problem, where the sequence of time slots represents the action space. In this way, the scheduler 1010 can determine how many time slots to select to deliver the correct number of packets to the correct destination, and can select the time slots in a given order. This problem is described using a continuous action space, representing a stochastic game with a continuous action space, thereby avoiding a large number of possible actions. This problem is also solved by combining supervised learning with RL techniques. Thus, three components are described: the state, the label, and the loss function, as well as the neural network NN.

[0086] The state or input features represent the state of the environment in the most recent H superframes. The state is described as [O n ,TX], where Each element in the matrix corresponds to And the network scheduling information TX∈{0,1} H×S×C , where each element in the matrix corresponds to By using this representation, it is possible to indicate when an observation can be used to identify a pattern. t,c = 1, the node uses the spectrum in channel c at time t, which means that the observation May contain energy from nodes in the network and may give a false impression of the pattern to be recognized.

[0087] To further construct the tag, time slots are defined as usable, unusable, or unknown. Therefore, φ is defined as the noise threshold. Usable and unusable time slots are defined in Equations 13 and 14.

[0088]

[0089]

[0090] In other words, in Equation 13, a usable time slot is a time slot where the amount of detected energy is below the noise threshold φ or where a successful transmission has occurred. Otherwise, as described in Equation 14, a time slot is defined as unusable if it is not available and is not part of the overall schedule but may be part of the node receiving the schedule. Time slots in a superframe that are not set as usable or are unusable are marked as unknown or not marked at all.

[0091] The loss is further described in Equation 15:

[0092]

[0093] The cross entropy loss for the labels that can be used and those that cannot be used is calculated and divided by the number of slots that could have been updated. The loss for unknown slots will be zero, and the NN will attempt to preserve the output values for those slots. To ensure that every example used for training is as important as all other examples, the cross entropy is divided by the number of slots that could have been updated. This prevents the NN from finding a way to start using all slots and reduces the loss by marking as many slots as possible as unknown.

[0094] NN is used to optimize the prediction. As an illustrative embodiment, a NN comprising eight fully connected layers may be used, where for each layer neurons, where l is the number of layers between [1,8]. A swish activation function can be used on each layer, except for the final output layer, where a softmax activation is applied. In addition, since data generation, scheduling strategies, network topology, and the environment can vary dramatically, online learning is preferred. After every δ steps, a training step is performed, in which random samples are selected from the replay buffer. To optimize the weights in the NN, the RMSProp optimizer can be used. As an illustrative example, the following hyperparameter values can be used:

[0095]

[0096] Except for Figure 1 and Figure 2 The different topologies shown in FIGURE 1 also enable different communication methods between nodes 100-105 and 200-206. As illustrative examples, five methods can be distinguished. The conventional method is used in most systems. To achieve high throughput, the conventional method uses all idle time slots in internal scheduling via scheduler 1010. In other words, it uses only traditional MF-TDMA scheduling algorithms or CSMA strategies without causing conflicts within the own network. Secondly, the "no transmission" method is an extreme method in which the own network does not transmit any messages at all. This is an extreme case, but it illustrates the optimal performance of INCs 110-111 and 210-211. Fourth, the optimal method, when α = 0.5, is a hypothetical optimal solution to the problem. In the optimal method, the own network knows the behavior of the INC in advance, which is not possible in reality. Finally, in the Deep Q Network (DQN) method, each node in nodes 100-105 and 200-206 uses a DQN, where, as an illustrative example, a superframe size of 1 is used. Finally, the fifth method is the method addressed by the present invention, which is called the prediction method.

[0097] exist Figure 3 In, for Figure 1The topology of exemplifies the average total throughput F(α) 300 of the superframe. It is observed that in the non-transmitting method, the total throughput decreases as α 301 increases. On the other hand, in the conventional method, the total throughput increases as α 301 increases. The non-transmitting method and the conventional method are the two most extreme methods used to indicate the performance of the INC and its own network. In addition, as shown in the figure, the DQN method performs poorly, while the performance of the prediction method of the present invention is at least as good as the conventional method when α < 0.8, or at least as good as the best method when α > 0.8.

[0098] In addition, Figure 4 , different network metrics for superframes in a MF-TDMA mesh network for different scheduling methods are illustrated. 405 illustrates that the prediction method of the present invention reduces the average number of collisions per superframe by 30% compared to the conventional scheduling method. Additionally, throughput is illustrated in 400, message success rate is illustrated in 401, missed opportunities are illustrated in 402, noise throughput is illustrated in 403, and INC message success rate is illustrated in 404. It is observed that the prediction method of the present invention is always between the performance of the conventional method and the performance of the optimal method. Since the prediction method is less aggressive than the conventional method, the throughput is reduced. If the module predicts a false positive, this will result in fewer collisions.

[0099] exist Figure 5 In, for Figure 2 The topology of exemplifies the average total throughput F(α) of the superframe 300. In this topology, when α < 0.7, the prediction method outperforms the conventional method and the number of collisions is reduced by a factor of 1.3, which is Figure 6 Reference 605 is shown. Figure 6 Illustrated in Figure 2 The network metrics of the superframes in the MF-TDMA AP network for different scheduling methods are shown in FIG. The throughput 600 of the own network is similar, and the throughput 600 of INC increases by 1.17 times, as shown in FIG. Figure 6 , the metric message success rate 601, missed opportunities 602, INC throughput 603 and INC message success rate 604 are further illustrated.

[0100] exist Figure 7 In , the average total throughput F(α)300 of a superframe is illustrated for the CSMA AP topology. Figure 8 The network metrics associated with this topology are illustrated in .

[0101] Figure 9A suitable computing system 900 is shown for implementing an embodiment of a method for predicting radio spectrum usage in an environment according to the present invention. Computing system 900 can generally be implemented as a suitable general-purpose computer and include a bus 910, a processor 902, a local memory 904, one or more optional input interfaces 914, one or more optional output interfaces 916, a communication interface 912, a storage element interface 906, and one or more storage elements 908. Bus 910 may include one or more conductors that allow communication between components of computing system 900. Processor 902 may include any type of conventional processor or microprocessor that interprets and executes programmed instructions. Local memory 904 may include random access memory (RAM) or another type of dynamic storage device that stores information and instructions for execution by processor 902, and / or read-only memory (ROM) or another type of static storage device that stores static information and instructions for use by processor 902. Input interface 914 may include one or more conventional mechanisms that allow an operator or user to input information into computing device 900, such as a keyboard 920, a mouse 930, a pen, voice recognition and / or biometric mechanisms, a camera, and the like. Output interface 916 can include one or more conventional mechanisms for outputting information to an operator or user, such as a display 940. Communication interface 912 can include any transceiver-like mechanism, such as, for example, one or more Ethernet interfaces that enable computing system 900 to communicate with other devices and / or systems, such as with one or more other nodes in node 100-105 or 200-206. The communication interface 912 of computing system 900 can be connected to another computing system via a local area network (LAN) or wide area network (WAN) (such as, for example, the Internet). Storage element interface 906 can include a storage interface (such as, for example, a serial advanced technology attachment (SATA) interface or a small computer system interface (SCSI)) for connecting bus 910 to one or more storage elements 908, such as one or more local disks, such as SATA disk drives, and controlling the reading and writing of data to and / or from these storage elements 908. Although the above (one or more) storage elements 908 are described as local disks, generally any other suitable computer-readable medium, such as a removable disk, an optical storage medium such as a CD or DVD, a ROM disk, a solid-state drive, a flash memory card, etc. can be used. The computing system 900 may therefore correspond to Figure 1 or Figure 2 Nodes in the embodiment shown in .

[0102] As used in this application, the term "circuitry" may refer to one or more or all of the following:

[0103] (a) hardware circuit implementations only, such as implementations in analog and / or digital circuitry only, and

[0104] (b) a combination of hardware circuitry and software such as (as applicable):

[0105] (i) a combination of (one or more) analog and / or digital hardware circuits and software / firmware, and

[0106] (ii) any portion of a hardware processor(s) with software (including digital signal processor(s), software and memory(s) that work together to enable a device such as a mobile phone or server to perform various functions), and

[0107] (c) Hardware circuit(s) and / or processor(s) (such as microprocessor(s) or portion(s) of microprocessor(s)) that require software (e.g., firmware) to operate, but the software may not be present when not needed for operation.

[0108] This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term "circuitry" also covers implementations that are solely a hardware circuit or processor (or multiple processors) or a portion of a hardware circuit or processor and its accompanying software and / or firmware. The term "circuitry" also covers, for example, and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or similar integrated circuit in a server, cellular network device, or other computing or networking device.

[0109] Although the present invention has been described with reference to specific embodiments, it will be apparent to those skilled in the art that the present invention is not limited to the details of the foregoing illustrative embodiments and that the present invention may be implemented with various changes and modifications without departing from the scope thereof. The embodiments given are therefore to be considered in all respects as illustrative and not restrictive, and the scope of the invention is indicated by the appended claims rather than the foregoing description, and all changes falling within the scope of the claims are intended to be covered thereby.

[0110] In addition, readers of this patent application will understand that the words "comprise" or "include" do not exclude other elements or steps, the words "a" or "an" do not exclude a plurality, and a single element such as a computer system, a processor or another integrated unit can perform the functions of several components set forth in the claims. Any figure marks in the claims should not be interpreted as limiting the corresponding claims involved. When used in the specification or in the claims, the terms "first", "second", "third", "a", "b", "c", etc. are introduced to distinguish similar elements or steps and do not necessarily describe a sequential or chronological order. Similarly, the terms "top", "bottom", "above", "below", etc. are introduced for descriptive purposes and do not necessarily indicate relative positions. It should be understood that, where appropriate, the terms used in this way are interchangeable and that embodiments of the present invention can operate in accordance with the present invention in other orders or in orientations different from those described or illustrated above.

Claims

1. A computer-implemented method for training a neural network for predicting radio spectrum usage in an environment comprising interference (110-111, 210-211) and a plurality of wireless nodes (100-105, 200-206), the wireless nodes (100-105, 200-206) operating within the radio spectrum via a slotted medium access control (MAC) protocol, the method comprising the steps of: - obtaining usage of each time slot of the radio spectrum by sensing the radio spectrum; -Mark time slots: Marking a timeslot as usable when the corresponding radio spectrum usage falls below a predefined noise threshold or when a wireless node has successfully transmitted a message through that timeslot; o Marking a timeslot as unusable when the timeslot is not marked as usable but the timeslot is suitable for receiving a message according to the MAC protocol; and For prediction, the loss function is calculated as the cross entropy loss of usable and unusable time slots divided by the total number of usable and unusable time slots; as well as - Train a neural network to predict available and unavailable time slots by minimizing a loss function.

2. The computer-implemented method of claim 1 , further comprising the steps of: -Optimize the layers of a neural network.

3. The computer-implemented method of claim 2 , wherein the optimizing further comprises: -Optimize the output layer of the neural network through softmax activation.

4. The computer-implemented method according to any one of claims 1 to 3, further comprising the steps of: -Prediction of upcoming radio spectrum usage of interference (110-111, 210-211) through neural networks.

5. The computer-implemented method according to any one of claims 1 to 3, wherein the slotted MAC protocol is a time division multiple access (TDMA) protocol, a multi-frequency time division multiple access (MF-TDMA) protocol or a slotted carrier sense multiple access (CSMA) protocol.

6. The computer-implemented method of any one of claims 1 to 3, wherein the sensing is performed by a software defined radio (SDR) platform.

7. A data processing system (900) comprising means for performing the method according to any one of claims 1 to 6.

8. A computer program product comprising computer executable instructions for executing the method according to any one of claims 1 to 6 when the computer executable instructions are run on a computer. 9 . A computer-readable storage medium storing computer-executable instructions for executing the method according to claim 1 when the computer-executable instructions are executed on a computer.

10. A wireless node (100-105, 200-206), comprising: a processor configured to operate via a slotted medium access control (MAC) protocol within a radio spectrum; The processor is further configured to predict available time slots by means of a neural network trained according to the computer-implemented method according to any one of claims 1 to 6.

Images