Vehicular cache and data forecast methodoligies using personal area networks

Abstract

Examples of systems described herein include vehicles having cache memory. Example systems may form a personal area network comprising the vehicle and one or more user devices. The personal area network may be formed using a wireless communication protocol. An example system may forecast certain data to be cached into the cache memory, based on activity in the personal area network. The example system may cache the certain data over the personal area network from the one or more user devices into the cache memory of the vehicle.

Description

Docket No. 0077145-07501VEHICULAR CACHE AND DATA FORECAST METHODOLIGIES USING PERSONAL AREA NETWORKSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit under 35 U. S. C. § 119 of the earlier filing date of U. S. Provisional Application Serial No. 63 / 725,378 filed November 26, 2024, the entire contents of which are hereby incorporated by reference in their entirety' for any purpose.TECHNICAL FIELD

[0002] Examples described herein relate generally to systems including cache memory in vehicles. Examples of forecasting data for storage in cache memory of a vehicle based on personal area network activity are described.BACKGROUND

[0003] Caching can be used for data management in vehicular networks such as vehicle-to-everything (V2X), vehicle-to-vehicle (V2V), and vehicle-to-infrastructure (V2I) systems. For instance, in V2X, caching optimizes data access and contributes to the seamless flow of information when vehicles interact with infrastructure elements like traffic signals. In V2V, caching may facilitate data exchange between vehicles to enhance communication efficiency and support advanced driver-assistance systems. In loV, where vehicles, infrastructure, devices, and users are seamlessly interconnected, caching may alleviate network congestion and optimize resource utilization.

[0004] Various caching solutions exist for vehicular networks. This includes federated deep reinforcement learning based on vehicle preferences and historical request frequency. Additionally, fog computing-based caching mechanisms have been developed for autonomous vehicle path planning in congested metropolitan areas. Edge vehicular caching schemes for vehicle paths leverage sparrow search optimization to reduce latency and improve hit rates. The cooperative caching utilizes Hawkes processes and deep reinforcement learning to optimize throughput and energy' consumption. Geographic distribution-aware caching strategies select optimal cache nodes among vehicles and roadside units to enhance energy efficiency. Mobility-aware solutions in incorporate neural networks for prediction, clustering, and security measures, while other approaches leverage long short-term memory (LSTM) networks to predict content requests for caching decisions. Furthermore, social-aware cooperative caching applies federated learning to reduce contentDocket No. 0077145-07501transmission latency, and edge collaborative content caching provides low-latency services for vehicles. Spatial-temporal vehicle prediction is applied for optimal content deployment and scheduling.

[0005] Despite these advancements, current approaches remain focused on V2V or V2I connectivity paradigms that focus the on connectivity of multiple vehicles in a centralized or decentralized hierarchy. This compels users to be heavily reliant on network operators, other vehicles, and infrastructure. This creates challenges including network and control overheads, limited control over cache placement, and substantial dependency on centralized data sources, which impedes the agility of vehicles within these systems. These paradigms hinder users from fully operating their vehicles autonomously within their own decentralized personal area network.SUMMARY

[0006] Example systems are disclosed herein. In an embodiment, an example system includes cache memory in a vehicle, at least one processor in the vehicle, and at least one computer readable media encoded with executable instructions. When executed by the at least one processor, the executable instructions cause the system to perform operations including form a personal area network comprising the vehicle and one or more user devices, the personal area network formed using a wireless communication protocol, forecast certain data to be cached into the cache memory, based on activity in the personal area network; and cache the certain data over the personal area network from the one or more user devices into the cache memory of the vehicle.

[0007] Additionally, or alternatively, the operations further include select the certain data from the personal area network to be cached into the cache memory. The operation select the certain data includes selecting most recently used data (mru), selecting least recently used data (Iru), selecting earliest-deadline first data (erf), selecting latest-deadline first data (Idf), selecting high-processing first data (HPF), or combinations thereof.

[0008] Additionally, or alternatively, the certain data originates from a source, the source including one or more of an access point, a cloud core, a mobile station, or the one or more user devices.

[0009] Additionally, or alternatively, the operations further include preload the certain data from the source into the vehicle prior to the operation form the personal area network.Docket No. 0077145-07501

[0010] Additionally, or alternatively, the operation forecast the certain data includes generate future demand data; and cache the future demand data over the personal area network from the one or more user devices into the cache memory of the vehicle.

[0011] Additionally, or alternatively, the activity in the personal area network comprises application usage generated by the one or more user devices.

[0012] Additionally, or alternatively, the operation generate the future demand data includes receive a request that is sent from the one or more user devices. The request comprises a request for the certain data being used by least one user device of the one or more user devices at a time period, timestamp the request with the time period to create a timestamped request, send the timestamped request to the cache memory. If the timestamped request does not satisfy a condition set forth by a subset state of the cache memory, instruct generation of an age counter for each incremented subset state of a plurality of incremented subset states. The each incremented subset state includes the subset state incremented by an integer, instruct generation of a demand forecast parameter. The demand forecast parameter includes a predicted demand for a subsequent request at a subsequent time period, and cache the demand forecast parameter into the cache memory of the vehicle.

[0013] Additionally, or alternatively, the condition includes a presence or an absence of the certain data requested for in the request.

[0014] Additionally, or alternatively, the integer is five minutes.

[0015] Additionally, or alternatively, the subsequent time period is one hour after the time period.

[0016] Additionally, or alternatively, the timestamped request is a first timestamped request, and the operations further include receive a second timestamped request from the one or more user devices, insert the second timestamped request into the cache memory, and in response to the inserting the second timestamped request, evict a non-compliant timestamped request that does not satisfy an eviction parameter. The eviction parameter includes the demand forecast parameter and the age counter.

[0017] Additionally, or alternatively, the personal area network is isolated to the vehicle, the one or more user devices, and the wireless communication protocol.

[0018] Additionally, or alternatively, the personal area network is isolated to the vehicle and the one or more user devices.

[0019] Additionally, or alternatively, the operations further include authenticate the one or more user devices to generate an authenticated one or more user devices. The operationDocket No. 0077145-07501authenticate includes input a command for the one or more user devices to provide an authentication parameter to the personal area network; and upon a successful authentication, the successful authentication includes the one or more user devices satisfying the authentication parameter, access the personal area network with the authenticated one or more user devices.

[0020] Additionally, or alternatively, the personal area network is isolated from the one or more user devices that have had an unsuccessful authentication. The unsuccessful authentication includes the one or more user devices failing the authentication parameter.

[0021] Additionally, or alternatively, the system further includes at least one antenna coupled to the vehicle. The at least one antenna is configured for wireless communication with the one or more user devices to form the personal area network.

[0022] Additionally, or alternatively, the at least one antenna communicates with the one or more user devices at a bandwidth between 6-60 GHz.

[0023] Additionally, or alternatively, the system further includes an array, the array includes one or more antennae. The one or more antennae communicates with the one or more user devices at a bandwidth between 6-60 GHz.

[0024] Additionally, or alternatively, the personal area network is isolated from the cloud core.

[0025] Additionally, or alternatively, the cloud core includes one or more of an edge computing server, cloud server, or fog computing server.

[0026] Additionally, or alternatively, the personal area network is isolated from the access point.

[0027] Additionally, or alternatively, the personal area network is isolated from the mobile station.

[0028] Additionally, or alternatively, the vehicle includes a power source configured to power the at least one processor.

[0029] Additionally, or alternatively, the operations further include receive a future request for the certain data that is sent from the one or more other user devices; and service the future request from the cache memory.

[0030] Example methods are disclosed herein. In an embodiment, an example method includes forming, using a computing device comprising a memory' and at least one processor, a personal area network including a vehicle and one or more other user devices, the personal area network formed using a wireless communication protocol. The methodDocket No. 0077145-07501may include forecasting, using the computing device, certain data from the personal area network to be cached into a cache memory of the vehicle. The method may include caching, using the computing device, the certain data over the personal area network from the one or more user devices into a cache memory of the vehicle.

[0031] Additionally, or alternatively, the forecasting includes generating, using the computing device, future demand data, wherein the future demand data includes a predicted demand for a future request at a subsequent time period. The subsequent time period occurs later than an initial time period.

[0032] Additionally, or alternatively, the generating the future demand data includes receiving, using the computing device, a request for the certain data that is sent from one or more user devices, the request including a timestamp, sending, using the computing device, the request to the cache memory; and generating, using the computing device, a plurality of counters for a corresponding plurality of future subset states of a cache memory. Each future subset state of the corresponding plurality of future subset states includes an initial subset state of the cache memory incremented by an integer.

[0033] Additionally, or alternatively, the method further includes isolating, using the computing device, the personal area network to the vehicle and the one or more user devices, the isolating includes severing, using the computing device, a connection between the personal area network and a source, the source comprising one or more of an access point, a cloud core, a mobile station, or one or more unauthorized user devices.

[0034] Additionally, or alternatively, prior to the isolating the personal area network to the vehicle and the one or more user devices, further including offloading, using the computing device, the future demand data onto the vehicle.

[0035] Example methods for training a machine learning model to generate a forecast model. In an embodiment, an example method of training a machine learning model to generate a forecast model includes receiving, using a computing device comprising a memory and at least one processor, a plurality of request vectors, each request vector of the plurality of request vectors comprising a request vector for an application at a time period. The method may include decomposing, using the computing device, the plurality of request vectors, to generate a plurality of trend vectors, and a plurality of seasonal vectors. The method may include passing, using the computing device, the plurality of seasonal vectors through a self-attention block, the self-attention block identifies correlation scores, producing a plurality of forecast vectors. The method may include mapping, using the computing device, the plurality of forecast vectors to a plurality of predicted demandDocket No. 0077145-07501vectors. The method may include generating, using the computing device, the forecast model by comparing the plurality of predicted demand vectors to a plurality of ground-truth aggregated demand vectors.

[0036] Additionally, or alternatively, the receiving includes receiving the plurality of request vectors from a user device connected to a personal area network.

[0037] Additionally, or alternatively, the method further includes applying, using the computing device, the forecast model to a cache memory. The cache memory assigns a priority rating to a plurality of future request vectors received by the computing device, based on the plurality of request vectors.

[0038] Additionally, or alternatively, the plurality of request vectors includes data from one or more of the following: most recently used data (mru), least recently used data (Iru), earliest-deadline first data (erf), latest-deadline first data (Idf), or high-processing first data (HPF).

[0039] Additionally, or alternatively, the comparing the plurality of predicted demand vectors to the plurality of ground-truth aggregated demand vectors includes minimizing an error between the plurality of predicted demand vectors and the plurality of ground-truth aggregated demand vectors over a horizon.BRIEF DESCRIPTION OF THE DRAWINGS

[0040] FIG. 1 is a schematic illustration of a system for a personal area network arranged in accordance with examples described herein.

[0041] FIG. 2 is a schematic illustration of a system for different phases of a personal area network arranged in accordance with examples described herein.

[0042] FIG. 3 is a flowchart of an example caching methodology arranged in accordance with examples described herein.

[0043] FIG. 4 is a flowchart of operation of an example machine learning training architecture arranged in accordance with examples described herein.

[0044] FIGS, 5 A and 5B are sheets of Equations referred to herein,

[0045] FIG. 6 is a flowchart of a method including forming a personal area network in accordance with examples described herein.

[0046] FIG. 7 is a flowchart of a method of training a machine learning model in accordance with examples described herein.

[0047] FIG. 8A illustrates a high cache cumulative hit in accordance with examples described herein.Docket No. 0077145-07501

[0048] FIG. 8B illustrates an efficiency of a proposed caching paradigm, in accordance with examples described herein.

[0049] FIG. 9 A illustrates an application of the system and method across different usage patterns, in accordance with examples described herein.

[0050] FIG. 9B illustrates observed validation trajectories across diverse applications in accordance with examples described herein.

[0051] FIG. 9C illustrates a hit rate in accordance with examples described herein.

[0052] FIG. 9D illustrates an efficiency across popularity quartiles in accordance with examples described herein.

[0053] FIG. 9E illustrates a correlation distribution analysis in accordance with examples described herein.DETAILED DESCRIPTION

[0054] Certain details are set forth herein to provide an understanding of described embodiments of technology. However, other examples may be practiced without various of these particular details. In some instances, well-known circuits, control signals, timing protocols, automotive components, networking components, machine learning or artificial intelligence components or techniques, and / or software operations have not been shown in detail in order to avoid unnecessarily obscuring the described embodiments. Other embodiments may be utilized, and other changes may be made, without departing from the spirit or scope of the subject matter and / or claims presented here.

[0055] Personal vehicles may be positioned as user-centric micro-servers that operate independently within the personal area networks. The use of personal vehicles in this way may promote and / or achieve data sovereignty as opposed to reliance on centralized infrastructures that dominate current automotive ecosystems. This micro-server (which may be exclusively allocated to their owners) may act as the cache, memory, and processing node that remain proximity to users throughout their daily activities (e.g., particularly in the United States). This proximity may allow partially and / or wholly seamless data synchronization and access. To allocate user data on this micro-server, caching may be used m advance based on user traffic. For example, reactive caching policies are often used that respond to access patterns only after they occur. However, this yields cache misses when users operate beyond home network boundaries, such as during transitions between connectivity zones where timely data access becomes useful for maintaining seamless userDocket No. 0077145-07501experience. To address this m some examples, predictive caching can be applied to anticipate future application needs prior to connectivity loss.

[0056] Learning approaches may face limitations m vehicular contexts. For example, recurrent neural networks (RNN) may suffer from vanishing gradients when modeling multi-day usage patterns, thus failing to capture the weekly periodicities inherent in commute-based application access. Further, convolutional neural networks (CNN) may lack the temporal capabilities used by time series forecasting and cannot generally decompose the complex mixture of trend and seasonal components present in vehicular usage data.

[0057] Examples of systems described herein include vehicles having cache memory. Example systems may form a personal area network comprising the vehicle and one or more user devices. The personal area network may be formed using a wireless communication protocol. An example system may forecast certain data to be cached into the cache memory, based on activity m the personal area network. The example system may cache the certain data over the personal area network from the one or more user devices into the cache memory' of the vehicle.

[0058] Examples described herein include systems operating in accordance with a user- centric caching paradigm that is focused on the user, independent from external entities such as other vehicles or infrastructure (e.g., cloud service providers). Examples described herein integrate personal vehicles into the user's personal area network. This paradigm transforms caching from centralized hierarchies to a decentralized model in personal vehicles, without connectivity to other vehicles. E ery user carries their own individual caches in their vehicles, thus empowering them to take control of their data content.Individuals often park their vehicles at their homes, near offices, shopping areas, and other frequently visited locations. This proximity of users to their vehicles can be leveraged by using personal vehicles as micro-servers and connectivity' hotspots in an extended personal area network model. The vehicle serves as a personalized cache carrying the user data continuously while offloading user data from centralized network servers such as cloud or edge / fog networks. Vehicles may be dynamic data centers that create, connect to, and / or participate in personalized networks that seamlessly integrate phones, Wi-Fi devices, and office networks. This approach allows users to store and access their data locally, hence reducing dependency on external network providers. Vehicles maybe equipped with phased arrays to support high data rates and extended connectivity ranges when needed, e.g., communicating with Wi-Fi networks, mobile networks, and other networks associated with the user data.Docket No. 0077145-07501

[0059] Examples of neural network architectures are described herein for personalized vehicular cache prediction. Examples employ prediction with tuned parameters optimized for slot granularity (e.g., five-minute slots in some examples). Examples of cache systems may utilize lookahead capability for proactive cache management (e.g., one-hour lookahead in some examples). Example architectures may integrate predictions with an age-aware eviction mechanism that balances forecasts against cache staleness, which eliminates and / or reduces separate prediction and eviction modules, e.g., reducing complexity while improving cache coherence.

[0060] Examples of systems described herein accordingly differ from examples of systems employing loV and V2V communication, despite sharing elements of connectivity within the automotive context in some examples. Examples of systems described herein may conduct personal data management (PDM) within individual vehicles. Examples utilize storage, retrieval, and local processing of a user personal information, e.g., creating a personalized area network that extends beyond the vehicle to include phones, home, and / or office networks. In contrast, loV may generally utilize around broader connectivity that involves interactions with infrastructure, cloud-based services, and other vehicles. In examples described herein, the vehicle serves as a decentralized data center for the user, offering a local cache for frequently accessed data. This stands in contrast to examples of loV, where systems leverage a centralized cloud or network infrastructure for data storage, processing, and sharing among various entities. Moreover, examples of systems described herein may be user-centric, in that the systems provide personalized and localized data services for individual vehicle owners. In this manner, users may manage and access their data within their personal vehicles. Meanwhile, loV encompasses a broader ecosystem that involves interactions between vehicles, infrastructure, and external services with a focus on enhancing overall traffic management, safety, and efficiency. Examples of systems described herein may not contradict with V2V, as communication may be provided between vehicles, such as to facilitate communications for road safety and transportation, contingent upon user preference and decision. In some examples, a user may offer leasing of their vehicle cache to other vehicles for a fee.

[0061] Examples of systems and techniques described herein may select data used in a personal area network and store the selected data in the cache of a vehicle. In this manner, accessing certain data from devices in or near the vehicle may be accelerated. In this manner, examples of systems described herein utilize a user-centric caching paradigm.

[0062] FIG. 1 is a schematic illustration of a system 100 arranged in accordance with examples described herein. The system of FIG. 1 includes vehicle 102, access point 118 andDocket No. 0077145-07501user device 116. The access point 118 and user device 116 are included in wireless LAN 120. The vehicle 102 forms a personal area network 122 with all or a portion of the wireless LAN 120. The vehicle 102 includes processor(s) 104, cache memory 108, antenna(s) 106, and computer readable media 110. The computer readable media 110 includes executable instructions for forming PAN 124, executable instructions for caching 112, executable instructions for retrieval from cache 114, and executable instructions for authenticating 126.

[0063] The components shown in FIG. 1 are exemplary. Additional, fewer, and / or different components may be used in other examples.

[0064] Examples of access point 118 described herein include, but are not limited to, houses, townhomes, condominiums, skyscrapers, warehouses, towers, huts, yurts, cabins, campers, recreational vehicles (RVs), tents, buoys, lighthouses, and offshore drilling stations.

[0065] Examples of systems described herein may utilize vehicles to form a personal area network and / or to cache data from the personal area network, which personal area network may include and / or communicate with a wireless LAN (WLAN),

[0066] A vehicle generally refers to an apparatus that is designed for transporting people and / or goods. A vehicle typically includes features for mobility such as wheels or wings and is powered by an engine or some form of propulsion. Vehicles can be categorized into various types based on their purpose, such as automobiles, trucks, motorcycles, boats, and aircraft. Examples of vehicles include, but are not limited to, automobiles, bicycles, electric bicycles, scooters, motorcycles, airplanes, helicopters, drones, vans, trucks, trains, boats, and submarines. Advantageously, a vehicle may travel with a user during all or a portion of a user's day - e.g., the vehicle can be expected to remain m proximity to a user. For example, a user may be expected to have their vehicle parked at or near their home while the user is at home and / or asleep. The vehicle can be expected to be parked at or near a place of business when the user is at work. The vehicle can be expected to be parked at or near a business while the user is at the business. In this manner, utilizing the vehicle as a cache may advantageously allow the user to access the cache at many different times during the user's day.

[0067] Examples of vehicles described herein may include cache memory, such as cache memory 108 of FIG. 1. Cache generally refers to hardware and / or software that stores data, typically for quicker access. Cache may act as a buffer between a main memory and a processor, reducing the time needed to access data. Caches are typically smaller but faster than main memory' and may- be organized hierarchically, with multiple levels (e.g., LI, L2,Docket No. 0077145-07501and L3). Cache memory may help to minimize the time taken to fetch data from the main memory, which in turn increases the overall speed of the system. Accordingly, data stored in cache memory 108 may be a copy of data which may be stored in one or more other memories or storage described herein, such as other memories or storage of vehicle 102, devices on the wireless LAN 120, and / or devices accessible through the Internet or other computing system. Cache may be implemented using memory circuitry of various types, including but not limited to volatile memory device(s) and / or circuitry, and / or circuitry integral with one or more processors. In one or more embodiments, a cache such as cache memory' 108 has a capacity' of C.

[0068] Examples of vehicles described herein may include one or more processors, such as processor(s) 104 of FIG. 1. Generally, any number or type of processor may be used including, but not limited to, one or more central computing units (CPUs), graphical computing units (GPUs), processor core(s), tensor core(s), programmable circuitry, controller, microcontroller, field programmable gate array (FPGA), and / or applicationspecific integrated circuitry (ASIC). In some examples, the processor(s) 104 and / or the cache memory' 108 of FIG. 1 may be implemented using a device, such as a Raspberry' PI device located in the vehicle, such as vehicle 102. The device can be disposed in many locations of the vehicle, such as the trunk, or any location that is adjacent to a power source of the vehicle. In some examples, the device can be removed from the vehicle.

[0069] Examples of vehicles described herein may include one or more computer readable media, such as computer readable media 110 of FIG. 1. The computer readable media 110 may be implemented using a variety of memory and / or storage devices including, but not limited to, read only memory' (ROM), random access memory' (RAM), solid state drives (SSD) or cards (SD cards), disk drives, or other memory and / or storage.

[0070] Examples of computer readable media described herein may be used to provide software (e.g., may store executable instructions, which, when executed by one or more processors, cause certain operations to occur). For example, computer readable media 110 may store executable instructions for forming PAN 124, executable instructions for caching 112, and / or executable instructions for retrieval from cache 114. While a single computer readable media 110 is shown in FIG. 1, any number may be used. While the various executable instructions are shown as being stored on a same computer readable media in FIG. 1, they may be stored on different computer readable media in some examples.

[0071] The executable instructions for forming PAN 124 may include instruction which, w'hen executed by the processor(s) 104, cause the vehicle 102 to form and / or join personalDocket No. 0077145-07501area network 122. The personal area network 122 may be formed by connecting the vehicle 102 to all or a portion of a wireless LAN, such as wireless LAN 120.

[0072] Accordingly, personal area networks may be formed in accordance with systems and techniques described herein. A personal area network (PAN) generally refers to a network designed for interconnecting devices associated with an individual, typically within a nearby distance range (e.g., a range of a few meters). A PAN accordingly primarily serves user devices of a user such as smartphones, laptops, tablets, wearable devices, desktops, servers, smart speakers, home automation devices, and / or appliances. PANs can be established through wired connections, like USB, and / or wirelessly using wireless communication protocols such as Bluetooth, Wi-Fi, and / or Zigbee. In some examples, the personal area network does not include or access devices not in a local area network associated with the personal area network. In some examples, the personal area network is isolated from a cloud core (e.g., one or more cloud service providers). In some examples, the personal area network is isolated from other vehicles.

[0073] Examples of personal area networks described herein, such as personal area network 122, may include a vehicle in communication with a WLAN, such as wireless LAN 120. A local area network (LAN) generally refers to a network that connects computers and devices within a limited area such as a home, school, or office building. It facilitates the sharing of resources like files, printers, and internet access among connected devices. LANs may utilize wired Ethernet connections, and / or wireless communication protocols such as Wi-Fi (e.g., a wireless LAN 120).

[0074] The executable instructions for caching 112 may include instructions to select certain data from the personal area network 122 and / or from the wireless LAN 120 to be cached. Any of a variety of caching techniques may be used including selecting most recently used data (mru), selecting least recently used data (Iru), selecting eariiest-deadline first data (erf), selecting latest-deadline first data (Idf), selecting highest-processing first data (HPF), or combinations thereof.

[0075] The executable instructions for caching 112 may further include instructions to cache the selected data in the cache memory of the vehicle, such as in cache memory 108 of vehicle 102 in some examples.

[0076] The executable instructions for retrieval from cache 114 may include instructions to retrieve data from the cache when subsequently requested. For example, the vehicle 102 and / or another user device (e.g., user device 116) may request data which has been stored inDocket No. 0077145-07501cache memory 108 the executable instructions for retrieval from cache 114 may retrieve the data from the cache memory 108 and provide it to the requesting user device.

[0077] In some examples, the executable instructions for forming PAN 124, executable instructions for caching 112, and / or executable instructions for retrieval from cache 114 may be referred to as a cache manager and / or as a cache management process.

[0078] Examples of vehicles such as vehicle 102 described herein may include one or more antennas, such as antenna(s) 106. The antenna(s) 106 may be in communication with processor(s) 104 and may be used for communication with one or more networks described herein, such as personal area network 122 and / or wireless LAN 120. The antenna(s) 106 can be extended from the interior of the vehicle 102 to the exterior of vehicle 102, or remain internal within the vehicle 102, In one or more embodiments, the antenna(s) 106 comprises a low-profile dual-band (2.4 / 5 GHz) multiple input, multiple output (MIMO) patch antenna assembly. Using a MIMO configuration allows the vehicle 102 to communicate with one or more of the multiple devices in the personal area networks 122, Furthermore, a MIMO assembly allows for enhanced throughput and download access speeds, in the event of the user's processing rate for using the certain data is high. In one or more embodiments, the antenna(s) 106 communicates with one or more user devices, such as user device 116, at a bandwidth between 6-60 GHz. In one or more embodiments, the antenna(s) 106 is part of an array, where at least one of the antenna(s) 106 communicates with the one or more user devices at a bandwidth of 6-60 GHz. In one embodiment, the array is a phased array. In one embodiment, the array is a uniform linear array (ULA). In one embodiment, the array is a uniform planar array (UP A).

[0079] In some examples, vehicle 102 described herein may have additional components. For example, vehicle 102 may have one or more power source(s) which may be used to power processor(s) 104 and / or antenna(s) 106. Power sources which may be used include, but are not limited to, one or more batteries and / or solar panels. In one or more embodiments, the one or more power source(s) will be the vehicle batten7and the alternator, with an auxiliary battery. In one embodiment, while the engine of the vehicle 102 is running, power may be provided from a 12V bus (e.g., the alternator) of the vehicle 102. This may be implemented by coupling the 12V accessory line through an automotive DC-DC regulator (e.g., wide-input, low-voltage cutoff, ignition-sense). This may allow for continuous function while the engine is running and / or for recharging an auxiliary7battery connected to the vehicle 102 while the vehicle 102 is being driven. In some examples, while the engine of the vehicle 102 is not running, the auxiliary battery provides power to the vehicle 102. The auxiliary battery may be disposed within the vehicle 102 via DC-DC railsDocket No. 0077145-07501behind a low-voltage disconnect, which may be to protect the auxiliary batten,'. In some examples, an ignition-sense controller manages the state of the engine of the vehicle 102, turning on and off the engine.

[0080] Accordingly, during operation the vehicle 102 may form a personal area network 122 including all or part of wireless LAN 120. Certain data stored and / or accessed by devices in the wireless LAN 120 may be cached by the vehicle 102. For example, frequently- accessed data by one or more devices in the wireless LAN 120 may be stored in cache memory 108. In some examples, data predicted to be accessed by- one or more devices in the wireless LAN 120 may be stored in the cache memory 108.

[0081] In this manner, selected data which had been in use by one or more devices in the wireless LAN 120 and / or which may be predicted to be used by one or more devices, maybe available m cache memory 108. Accordingly, as the vehicle 102 moves about and / or remains stationary near the wireless LAN 120, data may be requested by one or more user devices and / or by vehicle 102 and may be retrieved from the cache memory- 108. This may reduce and / or avoid a need to contact a cloud or external network, memory and / or storage latency to access selected data in some examples.

[0082] In some examples, the personal area network 122 described herein may have additional components. For example, the computer readable media 110 has executable instructions for authenticating 126. In some examples, upon receiving a request for the certain data from one or more user devices, such as the user device 116, the system may respond, in accordance with the executable instructions for authenticating 126 to the one or more user devices by inputting a command for the one or more user devices to respond with an authentication parameter. From there, the one or more user devices can either have a successful authentication or an unsuccessful authentication. If the one or more user devices satisfies the authentication parameter to the personal area network 122, then there is a successful authentication, creating an authenticated one or more user devices that connects to the personal area network 122. if the one or more user devices fails the authentication parameter, then there is an unsuccessful authentication, leading the personal area network 122 to isolate itself from the one or more user devices that have had an unsuccessful authentication. In some examples, the authentication parameter comprises one or more of: biometrics (e.g., eye / retinal scan, fingerprint), PIN / passcode / verification code, voice command, and facial recognition. In some examples, access to the personal area network 122 can be dual-mode, wherein access can be controlled from the one or more user devices and the vehicle 102. In one or embodiments, the authentication parameter includes a plurality of authentication parameters.Docket No. 0077145-07501

[0083] FIG. 2 illustrates a system 200 arranged in accordance with examples described herein. The system 200 may be present in multiple phases including: Phase IA 202, Phase IB 214, Phase II 220, and Phase III 224. The system 200 of FIG 2 may include a vehicle 208, mobile station 210, LAN 226, and / or cloud core 204. In some examples, the system 100 of FIG. 1 may be used to implement and / or may be implemented by the system 200 of FIG. 2. Accordingly, the system 100 of FIG. 1 may also experience connectivity phases as shown and described with reference to FIG, 2.

[0084] FIG. 2 is exemplary. Additional, fewer, and / or different components may be present m other examples. Additional, fewer, and / or different connectivity phases may be used in other examples.

[0085] In Phase IA 202, the personal area network includes a vehicle 208 (e.g., vehicle 102 of FIG. 1) in communication with a cloud core 204, an access point 206 (e.g., access point 118), and a mobile station 210. In some examples, the access point 206 includes a user device 216 (e.g., user device 116), In some examples, the vehicle 208 communicates with the cloud core 204, access point 206, and mobile station 210 via a LAN 226, such as the wireless LAN 120. In some examples, when the vehicle 208 is stationary at and / or near the access point 206, the system 200 may leverage the proximity of the vehicle 208 to the LAN 226 to perform offloading and / or synchronization of the data. For example, offloading and / or synchronization of the certain data can occur during off-peak periods of time. For example, Phase 1 A may occur when a vehicle is parked at or near a user's home. For example, when the vehicle 208 is parked in a driveway and / or garage it may have a connectivity pattern shown in Phase IA. Selected data may be cached, such as during overnight hours, such as while a user is sleeping.

[0086] In some examples, the certain data selected to be cached may include applications, software, and / or content. In some examples, the certain data may originate from a source. In some examples, the source may include one or more of an access point such as access point 118 or access point 206, a cloud core such as cloud core 204, a mobile station such as mobile station 210, or one or more user devices, such as user device 116 or user device 216.

[0087] In some examples, the cloud core 204 includes and / or may be implemented using one or more of an edge computing server, a cloud server, and a fog computing server. In this manner, data which is frequently used and / or predicted to be used by user devices may be cached in the vehicle 208. Such caching and / or synchronization may occur in some examples overnight.Docket No. 0077145-07501

[0088] Still referring to FIG. 2, the vehicle 208 (denoted as V), the access point 206 (denoted as AP), the cloud core 204 (denoted as CC), and the mobile station 210 (denoted as MS) may be interconnected by wireless or wired links that support cache synchronization, as shown by Eq. I in FIG. 5A. Each link £ (see Eq. 2 in FIG. 5 A) at time period t may be characterized by an available bandwidth B(£, t) and a propagation delay D(t, t), bounded by B(f ) and D(£), respectively.

[0089] Still referring to FIG. 2, a wireless communication interface, such as antenna(s) 106, may be used to implement various wireless protocols. A computer readable media, such as computer readable media 110, may include cache control logic that include request handling, forecasting, and eviction. Processors that access the computer readable media may interface with the cache (e.g., cache memory 108) and with the wireless communication interface via logical links. Power may be drawn from the vehicle’s 12V bus (e.g., regulated through a DC-DC converter) to supply these components with continuous and / or uninterrupted operations in some examples, e.g., during engine-off periods and electrical load fluctuations. The connectivity transitions among the three interfaces may be governed within the vehicle 208. For example, detection of the access point 206 service set identifier (SSID) on -V APmay trigger meshing and synchronization, along with the activation of {V:cc- Accordingly, the vehicle 208 is within the personal area network range. Given this proximity to the access point 206, data access may be satisfied. Henceforth, the cache contents may be preloaded as per the observed popularity, e.g., vehicle 208 may prefetch popular content from the access point, and reallocate long-term storage from the cloud core 204 to the vehicle 208. Thereafter, the vehicle 208 may partially and / or exclusively serve the mobile station 210 and / or user device 216 requests from its local storage (e.g., from cache memory).

[0090] Referring to Phase I A 202, in some examples, a user device 216 may send a command to the vehicle 208 and / or another trigger may occur. In some examples, the command may include one or more of selecting, transferring, or removing user content or applications from access point 206 and or user device 216 to the vehicle 208. In some examples, the command may include a command to migrate content from the cloud core 204 and / or the mobile station 210 to the vehicle 208. In some examples, the command may include a command to disconnect 212 from the cloud core 204 and the mobile station 210, which can occur before, during, or after offloading and / synchronization. In some examples, a disconnect 212 includes shutting off the communication between the vehicle 208 and the cloud core 204 and / or mobile station 210, and / or cancelling a subscription service to the cloud core 204 and / or the mobile station 210. This results in the personal area networkDocket No. 0077145-07501being isolated from the cloud core 204 and the mobile station 210. Furthermore, this may result in establishing the vehicle 208 as a consolidated, high-capacity data repository' for the user that is de-centralized from the cloud core 204, and the mobile station 210. For example, the vehicle's cache memory may be used to service one or more requests for data from one or more user devices in communication with the vehicle, even in situations where the vehicle and / or the user device are disconnected from a cloud core and / or a mobile station.

[0091] Referring phase IB 214, the system may be in a state where the personal area network includes the vehicle 208 being in communication with the LAN 226. In phase IB, the vehicle may be in communication with a personal area network, which may include LAN 226. The vehicle may be in communication with a user device, which may also be in communication with teh personal area network. The vehicle, however, may be disconnected from a cloud core and / or a mobile station. In some examples, the LAN 226 includes the access point 206, and / or the user device 216. In one or more embodiments, the cloud core 204 and / or the mobile station 210 are optionally in communication with the L AN 226, as illustrated by the dotted arrow lines. In some examples, the user device 216 may send a command to the vehicle 208 to disconnect 218 from the LAN 226. in some examples, another trigger event may occur to disconnect the vehicle from the LAN 226 or other component of a personal area network. For example, a particular time and / or date, or the vehicle may move a threshold distance from the LAN 226 or other component. In some examples, a disconnect 218 includes shutting off communication between the vehicle 208 and the LAN 226, or driving the vehicle 208 a distance from the LAN 226 until the LAN 226 is out of range. This results in the personal area network being isolated from the LAN 226. In one or more embodiments, the vehicle 208 may remain in communication with the user device 216 (e.g., the user device 216 may be in the vehicle and / or sufficiently near the vehicle to retain a communication link, such as Bluetooth or other wireless connection).

[0092] Phase II 220 depicts the system after disconnect 218. In the example of phase II 220, the personal area network includes the vehicle 208 and the user device 216. In some examples, the vehicle 208 is not in communication with the LAN 226. In this phase II, the vehicle 208 may retain (e.g., in cache memory ) recent data that the user accessed in previous sessions, such as previous sessions over LAN 226. In some examples, the vehicle 208 may store in cache memory data predicted to be requested by the user device 216 (e.g., based on application and / or other data activity in LAN 226), Accordingly, a user of user device 216 may seamlessly continue one or more sessions and / or engage in new sessions that are predicted based on usage patterns of previous sessions. For example, data that may be cached at the vehicle in PhaselA and / or Phase IB may be provided to the user device inDocket No. 0077145-07501Phase II. Various applications may be running on user device 216. The applications may receive data from LAN 226, cloud core 204, and / or mobile station 210 m Phase IA and / or IB, while in Phase II the applications may receive data from the cache of the vehicle 208. In some examples, the user device 216 may be in communication with the mobile station 210 in Phase II. In some examples, the vehicle 208 is not in communication with the mobile station 210 in Phase II. In some examples, the user device 216 may send a command to the vehicle 208 to connect 222 to the mobile station 210. In one or more embodiments, a connect 222 includes the vehicle 208 establishing communication with the mobile station 210 using an antenna, such as the antenna(s) 106 of FIG. 1. In some examples, the user device 216 ceases or has limited communication with the mobile station 210, due to the user device 216 being in a remote region or in an area with limited communication with the mobile station 210, or the user device 216 shutting off communication with the mobile station 210.

[0093] Phase III 224 depicts the system when the vehicle 208 is in communication with user device 216 and mobile station 210. In phase III 224, the personal area network includes the vehicle 208, the user device 216, and the mobile station 210. In some examples, the vehicle 208 is in communication with the user device 216 and the mobile station 210. In some examples, the user device 216 is in communication with the vehicle 208 and not the mobile station 210. In some examples, a multi-hop relay mechanism may be utilized to extend cached data access when users are at considerable remote distances from their cellular provider. For instance, if the user needs to conduct a live VoIP session, then the vehicle may serve as a data hub that maintains connectivity' through enhanced beam-forming architecture that may interconnect with a base station (e.g., mobile station 210), given the higher power capacity in the vehicle. In this manner, the vehicle 208 may be used to mediate communication between mobile station 210 and user device 216, For example, the vehicle may act as a relay node m emergency / disaster scenarios. Upon operation outside the vehicle 208-access access point 206 interface, e.g., vehicle 208 mobility, the vehicle 208-mobile station 210 interface may be reliant on the vehicle 208 content, given the proximity to the mobile station 210 that elevatesMSas the primary' link for low-latency service, while offloading the mobile station 210 from the cellular network. Therefore, in some examples, the vehicle 208 may act as a mobile micro-server traveling with the user, e.g., serving the user along the mobility path, before returning to the access point 206 home connectivity'. Throughout these operations, the system can remain self-contained. For example, module interactions that span request arrival through prediction and eviction, mayDocket No. 0077145-07501occur on the vehicles 208 platform, without reliance of external orchestration during distant data access and relay-connectivity.

[0094] FIG. 3 is a flowchart of an example caching methodology for caching or retrieving data, such as the certain data described with reference to FIG. 1, Accordingly, the system 100 of FIG. I and / or the system 200 of FIG. 2 may be used to implement the method shown and described with reference to FIG. 3. Although the example flowchart of FIG. 3 depicts a particular sequence of operations, the sequence may be altered in some examples without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or m a different sequence that does not materially affect the function of the architecture 300. In other examples, different components of an example device or system that implements the architecture 300 may perform functions at substantially the same time or in a specific sequence. In one or more embodiments, the flowchart of FIG. 3 may be implemented using executable instructions encoded in a computer readable media, such as the computer readable media 110.

[0095] According to some examples, the architecture 300 comprises a source 302, request counts 304, a vehicle request handler operation 306, a cache manager 308, a decision block310, a cache hit 312, a forecast operation 314, an eviction operation 316, an insertion operation 318, and an end stage 320,

[0096] Still referring to FIG. 3, in some examples, a cache, such as cache memory 108 of FIG. 1 may have a capacity C, the cache maintaining its state as a subset G of C application identifiers at time t (See Eq. 3 in FIG. 5A), where A denotes the total tracked applications (e.g., applications utilized by one or more user devices in a personal area network, such a the user device 116 of FIG. 1). The architecture 300 executes cache control logic that include request handling, forecasting, and eviction.

[0097] Still referring to FIG. 3, a source 302 sends request counts 304 modeled as sequential vectors (see Eq. 4 in FIG. 5A), where rt; is the count of requests for application i during the time period t. In one or more embodiments, the source 302 is a user device, such as the user device 116 or user device 216. The requests may be requests for data generated by one or more applications o the user device (e.g., requests for content such as Internet pages, text, media, videos, etc,). The vehicle request handler operation 306 may receive the request counts 304 and may timestamp each request rt,i with the time period t as a timestamped request. The vehicle request handler operation 306 may forward the newly created timestamped request to the cache manager 308. In some examples, the vehicle request handler operation 306 may be implemented using executable instructions, such asDocket No. 0077145-07501instructions encoded in the executable instructions for retrieval from cache 114 of FIG. 1. In one or more embodiments, the cache manager 308 may be implemented using the cache memory 108.

[0098] The architecture 300 may accordingly utilize a dataset that records requests per application in fixed time intervals. Examples of systems described herein may include executable instructions to observe a finite set of A applications whose request counts are recorded over T consecutive time periods. The A applications may be applications operating by devices within a personal area network (e.g., user devices). At slot t, the request vector rt,i, an A-dimensional row whose ith entry r® gives the requests for application i. The cache can hold up to C items. The cache's state at t is the set Ctof C application identifiers.

[0099] Moving to decision step 310, a cache hit 312 occurs if the timestamped request at index i matches or satisfies a condition set forth by a subset state Ct (i G Ct). In one or more embodiments, a condition comprises a presence or an absence of the certain data requested for m the timestamped request rt,t. In one or more embodiments, the certain data comprises data requested and / or utilized by one or more applications, and content. If the application and / or content requested by the user exists in the cache, then it is considered a cache hit. In one or more embodiments, if the application and / or content requested by the user does not exist in the cache, then it is considered a cache miss, and the operation proceeds to forecast operation 314 to generate future demand data.

[0100] To generate future demand data, the architecture 300 proceeds to forecast operation 314, where the cache manager 308 maintains an age counterfor each application i matching each incremented subset state Ct (as shown by Eqs. 5 and 6 in FIG. 5A) of a plurality of incremented subset states, incremented by one integer at each time period unless reset to zero upon access or insertion. In more detail, the integer is a size of each time period (also indicated as a time resolution) of the collected certain data. In some examples, the integer is a range of 1-5 minutes. At 1-2 minutes for the integer, the architecture 300 reacts faster in situations such as pre-fetching of the certain data or faster caching within short ranges between time periods. In some examples, the integer is five minutes. At five minutes for the integer, burst on-off behavior of the certain data is captured without excessive gaps between each request. Other time periods may be used in other examples.

[0101] Still referring to forecast operation 314, a plurality of demand forecast parameters may be generated, as shown by Eq. 7 in FIG. 5A, where each demand forecast parameter of the plurality of demand forecast parameters r̂t+1:t+H represents a predicted demand for a subsequent request for an application i at a subsequent time period (t + H), These demandDocket No. 0077145-07501forecast parameters may inform the cache replacement decision. The demand forecast parameter may be cached into the cache of the vehicle (e.g., cache memory 108 of FIG. 1 in some examples). In more detail, the subsequent time period is a lookahead duration variable based on the used prediction operation and the certain data. In some examples, the subsequent time period is a range of 1-3 hours. At 1-2 hours for the subsequent time period, the user accesses and uses the certain data in a volatile and / or burst manner, such as responding to a text, message, email, or performing a less data-intensive task. At 2-3 hours for the subsequent time period, the user accesses and uses the certain data in a stable and consistent manner, such as using a streaming content provider, viewing media, playing a game, or performing a more data-intensive task. In some examples, the subsequent time period is one hour after the time period. Other subsequent time periods may be used in other examples.

[0102] Moving to eviction operation 316, to inform evictions, for each cached application i matching Ct, its age counter aᵢ is defined as the number of integers elapsed since its last access. The age counter, as a vector at(shown by Eq. 5 in FIG. 5A), encodes recency information that the eviction operation 316 later combines with the demand forecast parameter. The age counters and the demand forecast parameters may be enqueued for consumption by the eviction operation 316, which evaluates each application i by computing a score function(See Eq. 8 in FIG. 5A). The eviction score for item i at time period t is given by Eq. 8 in FIG. 5A, where > 0 is a tunable age-penalty coefficient that may trade-off predicted popularity against staleness. In some examples, the timestamped request maybe a first timestamped request that may be followed by a second and / or other subsequent timestamped requests and / or requested data. In response to a miss for a second timestamped request from the source 302 for a new application j g Ct, the policy may select i* applications for eviction as non-compliant timestamped requests, as shown by Eq.9 of FIG. 5 A. An z* application may be a non-compliant timestamped request that does not satisfy an eviction parameter. In some examples, the eviction parameter may include the demand forecast parameter and the age counter.

[0103] Moving to insertion operation 318, the architecture 300 may instead insert the second timestamped request for the new application j (see Eq. 10 in FIG, 5 A), and increment ages of all remaining items by one integer. The methodology may integrate age-aware eviction in a neural architecture, which may avoid and / or reduce pitfalls of multi-stage pipelines. In some examples, the insertion operation 318 may be included in the executable instructions for caching 112 of FIG. 1. At end stage 320, the operation ends and the newDocket No. 0077145-07501applications and / or content may reside in the cache of the vehicle. In one or more embodiments, the architecture 300 receives a future request for data, such as the certain data, the future request being sent from the source 302. The architecture 300 may service the future request from the cache manager 308.

[0104] FIG. 4 is a flowchart of operation of training a machine learning model to generate a forecast model arranged in accordance with examples described herein. The flowchart 400 includes a source 402, a sliding window' tensor 404, an operation 406, parallel matrices 408, linear mapping 410, multi-headed self-attention block 412, matching and flattening operation 414, decoder mapper 416, and forecast model 418. The training of a machine learning model may be performed by one or more computing systems, such as computing systems described herein. In some examples, the vehicle 102 may perform the training, however, in some examples, a machine learning model may be trained by another computing system (not depicted in FIG. 1) and a trained model may be accessed and / or executed by the vehicle and / or other components in systems described herein. Accordingly, a computing system may be provided including one or more processors and computer readable media encoded with executable instructions for training a machine learning model to provide a forecast model. The computing system may perform the operations shown in FIG. 4. In some examples, circuitry may be used to perform the operations shown in FIG. 4 and / or implement the trained forecast model.

[0105] FIG. 4 is exemplary. Additional, fewer, and / or different operations may be used in other examples to provide a forecast model.

[0106] The training framework of FIG, 4 generally integrates series decomposition and auto-correlation mechanisms into a prediction-driven cache-replacement engine. Referring to FIG. 4, the flowchart 400 describes a sliding window tensor 404, as shown by Eq. 11 in FIG. 5 A, where W is the window length that captures the most recent history. Each row of the sliding window tensor 404 corresponds to the request vector for an application at a time period, in which this aggregated history may serve as a raw feature. In one or more embodiments, a source 402, such as the source 302, may provide data for the sliding window tensor 404 used by the flowchart 400.

[0107] Moving to operation 406, rather than feeding the raw counts directly, each request vector in the plurality of request vectors in the sliding window tensor 404 may first be normalized channel-wise using the maximum entry observed in the dataset. Thereafter, each column of the sliding window tensor 404 may undergo a moving-average decomposition of k, which may disentangle long-term shifts in user behavior from recurrent usage patterns, asDocket No. 0077145-07501shown by Eq. 12 in FIG. 5 A. This may result in parallel matrices 408 including a plurality of trend vectors Ttand a plurality of seasonal vectors Stof size W * A. The plurality of trend vectors Ttgenerally capture slow trends such as a gradual increase in streaming applications use over weeks, while the plurality of seasonal vectors Stmay contain high-frequency fluctuations. In one or more embodiments, as an additional step of operation 406, the sliding window tensor 404 is first linearly projected into a sequence of d-dimensional feature vectors, which produces the embedding function as shown by Eq. 16 in FIG. 5B. In more detail, the embed function in Eq. 16 is a front-door embedding. In one or more embodiments, the linear projection of the sliding window tensor 404 occurs before decomposition.

[0108] Moving to linear mapping 410, the plurality of seasonal vectors may be linearly embedded into a d-dimensional feature space via a learned projection, as shown by Eq. 13 in FIG. 5A. In some examples, systems utilizing the method of FIG. 4 may incorporate an encoder. In some examples, the encoder includes L stacked Layer modules. For each encoder layer 1 =1, 2,..., L, the plurality of seasonal vectors may be linearly embedded into a d-dimensional feature space via the learned projection, as shown by Eq. 13 in FIG. 5A. In one or more examples, the embed function in Eq, 13 in FIG. 5 A is an in-layer embedding.

[0109] Still referring to FIG. 4, the plurality of seasonal vectors may pass through a multi¬ headed self-attention block 412 that generates context- aw- are representations, which may emphasize recurring patterns across time. In some examples, the method of FIG. 4 may leverage a sparse auto-correlation mechanism in the multi-headed seif-attention block 412, For each head of the multi-headed self-attention block 412, the training method may identify correlation scores by identifying the top-k' lags with highest correlation scores between past and current embeddings, and attends to those positions. This may reduce the complexity from O(W2) to O(WlogW). The attention output may then be projected back into dimension d and the component i "ˡ is re-added in feature space. This yields the updated sequence Hˡ as shown by Eq. 14 in FIG. 5A, which may preserve seasonal detail and / or long-term shifts. In further detail, in Eq. 14 in FIG. 5A, attn denotes the standard transformer encoder layer, and proj denotes a point- wise linear layer. In one or more examples, Eq. 17 in FIG. 5B is utilized to produce the updated sequence Hˡ.

[0110] After L layers, the training methodology of flowchart 400 moves to the matching and flattening operation 414, where the final feature tensor Hˡ may be reduced by selecting its last H steps, for instance, matching forecast horizon and flattening to obtain a plurality of forecast vectors, shown by Ĥ.Docket No. 0077145-07501

[0111] Moving to decoder mapper 416, a fully connected decoder may map the plurality of forecast vectors to a plurality of predicted demand vectors, as depicted by Eq. 15 in FIG. 5A.

[0112] Moving to the forecast model 418, a forecast model may be generated by comparing the plurality of predicted demand vectors to a plurality of ground-truth aggregated demand vectors. In one or more embodiments, the network may be trained using a mean-squared error (MSE) on the plurality of predicted demand vectors compared against the plurality of ground-truth aggregated demand vectors over the next H horizon. The horizon is a lookahead timeframe based on the prediction of usage and the certain data. In an embodiment, the horizon is the subsequent time period. In some examples, an error, such as the MSE, between the plurality of predicted demand vectors and the plurality of groundtruth aggregated demand vectors may be minimized to generate the forecast model. The forecast model may accordingly be a trained machine learning model. During operation, forecast models described herein may receive user data from one or more personal area networks and generate a forecast of predicted data access(es). The predicted data access(es) may in some examples be used to select data to cache in a vehicle in systems described herein.

[0113] In some examples, the forecast model 418 may be utilized to manage cache described herein, such as cache memory 108 of FIG. 1. For example, the executable instructions for caching 112 may include instructions for accessing and / or utilizing a forecast model. In some examples, the computer readable media 110 of FIG. 1 may store data representing the forecast model, such as one or more weights and / or connections used m the model. In some examples, the system of FIG. 1, such as one or more processors operating in accordance with the executable instructions for caching 112 may assign a priority rating to a plurality' of future request vectors received by a computing device, such as the vehicle 102 or vehicle 208, that are based on the plurality of request vectors. The priority rating may be stored in the cache memory' 108 and / or in other electronic storage in communication with the vehicle 102. In some examples, the plurality of future request vectors are request vectors that are predicted to be sent and received at a future time period and / or date.

[0114] FIG. 6 is a flowchart of an example method 600 including forming a personal area network arranged in accordance with examples described herein. Although the example method 600 depicts a particular sequence of operations, the sequence may be altered without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or m a different sequence that does notDocket No. 0077145-07501materially affect the function of the method 600. In other examples, different components of an example device or system that implements the method 600 may perform functions at substantially the same time or in a specific sequence.

[0115] According to some examples, the method includes forming, using a computing device comprising a memory and at least one processor, a personal area network including a vehicle and one or more other user devices, the personal area network formed using a wireless communication protocol at block 602. In one or more embodiments, the personal area network may be implemented using the personal area network 122 of FIG. 1. In some examples, the vehicle may be implemented by vehicle 102 of FIG. 1 or vehicle 208 of FIG.2. In some examples, the one or more user devices may be implemented using the user device 116 of FIG. I and / or user device 216 of FIG. 2, The personal area network may be formed in some examples utilizing a processor executing executable instructions for forming PAN 124 of FIG. 1.

[0116] According to some examples, the method includes forecasting, using the computing device, certain data from the personal area network to be cached into a cache memory' of the vehicle at block 604. In one or more embodiments, forecasting includes performing all or portions of the methods shown and described with respect to FIG. 3and / or FIG. 4, including the forecast operation 314 in FIG. 3. In some examples, forecasting includes utilizing one or more machine learning models, such as a forecast model described herein. In some examples, the forecasting in block 604 includes generating, using the computing device, future demand data, wherein the future demand data includes a predicted demand for a future request at a subsequent time period, wherein the subsequent time period occurs later than an initial time period. In some examples, generating the future demand data comprises receiving, using the computing device, a request for the certain data that is sent from one or more user devices, the request comprising a timestamp. The computing device may send the request and / or data associated with the request, to the cache memory. The computing device may generate a plurality of counters for a corresponding plurality' of future subset states of a cache memory, wherein each future subset state of the corresponding plurality of future subset states comprises an initial subset state of the cache memory incremented by an integer. In some examples, the request is a request for the certain data being used by the one or more user devices. In some examples, the plurality of counters is implemented using and / or includes the age counter.

[0117] According to some examples, the method includes caching, using the computing device, the certain data over the personal area network from the one or more user devicesDocket No. 0077145-07501into a cache memory of the vehicle at block 606. In some examples, the cache memory may be implemented using the cache memory 108 of FIG. 1.

[0118] According to some examples, the method 600 further comprises isolating, using the computing device, the personal area network to the vehicle and the one or more user devices, the isolating comprises: severing, using the computing device, a connection between the personal area network and a source, the source comprising one or more of an access point, a cloud core, a mobile station, or one or more unauthorized user devices (e.g., as shown and described with reference to the phases of FIG. 2). In some examples, prior to isolating the personal area network to the vehicle and the one or more user devices, an extra step occurs where the computing device offloads the future demand data onto the vehicle (e.g., stored in a cache memory' of the vehicle).

[0119] FIG. 7 illustrates an example method 700 for training a machine learning model to generate a forecast model. Although the example method 700 depicts a particular sequence of operations, the sequence may be altered without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or in a different sequence that does not materially affect the function of the method 700. In other examples, different components of an example device or system that implements the method 700 may perform functions at substantially the same time or in a specific sequence.

[0120] According to some examples, the method 700 includes receiving, using a computing device comprising a memory and at least one processor, a plurality of request vectors, each request vector of the plurality of request vectors comprising a request vector for an application at a time period at block 702, In some examples, the computing device may be implemented using the vehicle 102 of FIG. 1, which may receive the plurality of request vectors from the source 402 (e.g., from one or more user devices, such as user device 116 and / or other devices in the personal area network, such as in the wireless LAN 120). In some examples, the source 402 may be a user device. In some examples, the plurality of request vectors include the sliding window tensor 404. In some examples, the user device may be connected to a personal area network, such as personal area network 122 of FIG. 1, In some examples, the plurality of request vectors includes data from one or more of the following: most recently used data (mru), least recently used data (Iru), earliest-deadline first data (erf), latest-deadline first data (Idf), or high-processing first data (HPF). In some examples, the data is the certain data.

[0121] According to some examples, the method 700 includes decomposing, using the computing device, the plurality of request vectors, to generate a plurality of trend vectors,Docket No. 0077145-07501and a plurality of seasonal vectors at block 704. In some examples, the decomposing of block 704 may include operation 406 of FIG. 4. In some examples, the plurality' of trend vectors and the plurality of seasonal vectors may include and / or may be implemented using the parallel matrices 408 of FIG. 4. In some examples, block 704 may include the linear mapping 410 function.

[0122] According to some examples, the method 700 includes passing, using the computing device, the plurality of seasonal vectors through a self-attention block. The self-attention block may identify correlation scores, producing a plurality of forecast vectors at block 706. In some examples, the self-attention block may include and / or may be implemented by the multi-headed self- ttention block 412 of FIG. 4. In some examples, block 706 may include the matching and flattening operation 414 of FIG, 4.

[0123] According to some examples, the method 700 includes mapping, using the computing device, the plurality of forecast vectors to a plurality of predicted demand vectors at block 708. In some examples, the mapping includes utilizing the decoder mapper 416 of FIG. 4.

[0124] According to some examples, the method 700 includes generating, using the computing device, the forecast model by comparing the plurality of predicted demand vectors to a plurality of ground-truth aggregated demand vectors at block 710. In some examples, the forecast model of block 710 includes and / or may be implemented using the forecast model 418 of FIG. 4. In some examples, the comparing operation includes calculating an error between the plurality of predicted demand vectors and the plurality of request vectors over a horizon. In some examples, an error meeting a condition (e.g., minimizing the error) may be used to generate the model.

[0125] According to some examples, the method 700 includes, using the computing device, operating the forecast model to manage a cache memory, including assigning a priority rating to a plurality of future request vectors received by the computing device, based on the plurality of request vectors. The priority rating may be stored in the cache memory in some examples. In some examples, the cache memory is the cache memory 108.

[0126] Accordingly, systems and methods for utilizing cache memory' of a vehicle are described herein. The cache memory' of the vehicle may store frequently accessed data and / or data predicted to be accessed by one or more user devices. The data stored in the cache memory of the vehicle may be selected in accordance with data utilized in a personalDocket No. 0077145-07501area network. The data stored in the cache memory of the vehicle may in some examples be selected using a forecast model.

[0127] Examples of systems described herein accordingly may provide multiple benefits for users and centralized networks. Examples of benefits and advantages are described herein to facilitate an appreciation for the described technology. It is to be understood that not all examples may impart alt or even any, of the described advantages. Advantages of some examples include low-latency access, reduced power consumption, higher capacity, improved spectral efficiency through data offloading and continuous connectivity', expanded coverage, enhanced content availability', cohesive end-to-end architecture, and strengthened user privacy.

[0128] For users, examples of systems described herein may provide enhanced data access speeds, as caching personal data in vehicles may facilitate faster access to frequently used information. Users can retrieve data from the local cache in their vehicles, reducing reliance on network providers and minimizing latency. Further, examples may strengthen connectivity in remote (poorly covered) areas with vehicles acting as data showers.Examples may reduce subscription costs by offloading data from personal centralized cloud storage to vehicles. Further, users rely on their personal cache for routine data needs, which may reduce and / or minimize their dependence on cellular networks. In some examples, caching in vehicles may give users more control over their data, thus reducing exposure to potential security' attacks and breaches associated with external networks. Examples may provide flexibility in PDM as users can prioritize certain types of data for caching, customize storage policies, and control access to their personal network. Examples may feature resilience to network failures, where the personal cache ensures and / or promotes continued access to certain data (e.g., essential data), thus reducing the impact of network outages,

[0129] Example benefits for network providers include reduced congestions as user reliance on vehicle caches alleviates the burden on its resources. Freed resources can be utilized for new users when traffic increases, hence enhancing network capacity. In some examples, as users rely more on local caches, operators may experience reduced pressure to expand their central infrastructure. In some examples, at low traffic density, servers can power down when few or no users are present in an area, which may yield less power consumption and higher energy efficiency.

[0130] Implemented ExampleDocket No. 0077145-07501

[0131] Examples of systems described herein may exhibit beneficial performance. A goal of example systems may be to predict incoming data needs and prefetch them to the vehicle, e.g., allowing the user to retrieve data directly from the vehicle’s cache. Examples of userspecific caching operate through real-time pattern recognition based on granular behavioral metrics. An example dataset was collected over an observation period across both individual and collective user environments. Empirical results demonstrated distinctive application utilization signatures with statistically significant consistency. For instance, User 1 heavily utilized applications A, B, D, and E with session durations ranking between 3500-3800 over data recorded over two weeks. Meanwhile, User 2 showed a preference for Applications A, B, C, and G with intervals spanning 4800-5450 minutes, e.g., a different utilization profile, Albeit with partial application overlap. The same distinct patterns were recorded for Users 3 to 5. Overall, recorded real-time data showed that there was a content popularity pattern across all users, e.g., 80%, 82%, 77%, 62%, and 64% of all usage time is for 57% of the applications for Users 1, 2, 3, 4, 5, respectively.

[0132] Moreover, temporal granularity analysis revealed additional stratification in application engagement patterns when examined at higher resolution. For example, User 3 activity was isolated across a 12-hour daily cycle, exposing distinct usage signatures. There were predominant temporal patterns throughout the day. For instance, morning hours showed predominant usage of Applications B and C, e.g., 63% of the usage time is for only 28% of applications. Meanwhile, evening hours demonstrated peaks for Applications B, C, and F that consumed 74% of usage time.

[0133] In collecting the aggregated frequency across all WLAN users with hourly tracking precision for all applications, it was shown that most applications consistently registered activity at frequency level 2, albeit varying usage time. Periodic spikes to levels 3 and 4 occurred throughout the timeframe with extended access time. Few applications reached maximum spikes at frequency level, particularly for Application A which has the highest access frequencies with 340 hours, followed by Application B. Overall, these empirical observations provided in the user behavior data demonstrated a significant pattern of concentrated application usage, which creates a foundation for effective caching strategies.

[0134] Caching may be performed via a set of dynamic approaches that shows a high cache cumulative hit that cached data was successfully retrieved, as shown by FIG. 8A. Namely, least-recently used (LRU), most-recently used (MSU), earliest deadline first (EDF), least storage first (LSF), and highest-processing first (HPF) caching methods are performed for the data collected, with a cumulative hit count that approaches 450-500 atDocket No. 0077145-075011000 request timestep, with high cache hit rate, that demonstrates the efficiency of the proposed caching paradigm, as shown by FIG. 8B.

[0135] A simulation of example systems described herein, such as an implementation of the methodologies shown and described with reference to flowchart 300 and flowchart 400 of FIG. 3 and FIG. 4, revealed predictive capabilities when applied to the system 100 and the system 200 by recording key performance metrics, which demonstrated the efficacy of the architectures. The configuration settings reflected a balance between model capacity' and computational efficiency. For example, an embedding dimension set at 122 provided sufficient representational capacity' in some examples for the application feature space while maintaining feasibility for real-time inference on automotive-grade processors, with each of the 4 attention heads operating on 32-dimensional subspaces that enable specialized pattern detection without excessive parameter overhead. The 2-layer encoder depth prioritized inference speed over depth, as testing revealed diminishing returns beyond two layers given the structured nature of usage patterns. The feed-forward dimension at 256 maintained the standard 2:1 expansion ratio from the model dimension. The kernel size of 25 corresponded to just over 2 hours of historical context, chosen to capture short-term periodicities while maintaining stability in the moving average computation. The learning rate configuration at 10"3with Adaptive Moment Estimation with Weight Decay (AdamW) leveraged adaptive estimation suitable for sparse gradients in usage prediction, while the weight decay of 10-4provided gentle L2-regularization that reduced and / or prevented parameter explosion without overly constraining the model's ability to learn sharp temporal transitions. The gradient clipping threshold at 1.0 facilitated training stability during early epochs when the random initialization may otherwise produce extreme gradients. The age-penalty coefficient of 10-2reduced and / or prevented cache staleness without overwhelming the predictive scores. The choice of batch size 50 training epochs was useful for the gradient stability and convergence. The application prediction analysis illuminates the model’s varying success across different usage pattern archetypes, as shown in FIG. 9A. Other parameters may be used in other examples.

[0136] A prediction for one application, which is referred to as AppA, which captures major spikes like a 30-unit peak at time step 190 with 80% accuracy. This demonstrates the effectiveness of the multi-head attention in identifying significant events. Moreover, a second application AppB, alignment-maintained predictions within the 5-25 range with correlation above 0.7. This exemplifies the model’s strength with quasi-periodic patterns that align with the architecture's decomposition. With a third application, App ground truthDocket No. 0077145-07501exhibited 0-25 fluctuations while predictions remained smoothed between 5-15 units. This reflected a tension between the tendency toward mean reversion and the need to capture extreme values. The MSE demonstrated the complexity of the prediction starting with 2.2 error for the first epochs. However, reduction to 0.15 was achieved at the final epochs. This generally validated the AdamW configuration with rate q0= 10". A scheduler exhibited two decay events at epochs 8 and 13 (reducing the learning rate by 0.5) which prevented premature convergence while enabling fine-grained optimization. A 1.0 clipping norm -was used for stable training, given the potential for exploding gradients in the deep attention layers. Divergence between training and validation curves, with Appcshowing validation MSE of 3.6 versus training MSE near 0.15, suggested mild overfitting that the weight decay partially but not completely mitigates.

[0137] The observed validation MSE trajectories in FIG. 9B demonstrated the example model’s generalization across the diverse applications. While App4and Appcshowed higher validation MSE reaching 3.6, this reflected the assessment of unpredictable applications rather than overfitting, as evidenced by consistent trajectories after epoch 20. A majority of applications in the implemented example (AppE, AppE, AppG, and AppB) maintained stable validation MSE below 2.5 throughout training, with AppGachieving exceptional performance near 1. This confirmed the model’s ability to capture regular usage patterns effectively. The average validation MSE stabilizes around 2.4 after epoch 30, which indicates proper convergence without significant overfitting despite the complex prediction. The early learning rate decay events at epochs 8 and 13 prevent the validation performance degradation, permitting stable refinement in later epochs.

[0138] The hit rate, as shown in FIG. 9C shows temporal stability with sporadic few misses occurring at irregular intervals such as steps 250, 350, and 600. These isolated events constitute less than 5% of operational time, which can correspond to cold-start scenarios for previously inactive applications. The rarity of these events validates the 12-integer forecast horizon, which provides a lookahead sufficient for preemptive population while avoiding overextension into uncertain future periods.

[0139] The performance demonstrates efficiency across all popularity quartiles as shown in FIG. 9D, with the cumulative hit rate curves exhibiting rapid convergence. The uniform achievement of 0.95 hit rate within 50-time steps across all quartiles validates the age¬ penalty coefficient selection, which effectively balances predictive scores against staleness indicators. The fastest convergence observed in Quartile 4 (reaching 0.90 hit rate within 20 steps) shows how the model quickly identifies high-frequency access patterns. Further, theDocket No. 0077145-07501cache capacity of C = 122 items proves adequate for the application set, as evidenced by the sustained 0.99 hit rate post-convergence, thereby indicating minimal capacity-induced evictions.

[0140] A correlation distribution analysis in FIG. 9E shows the accuracy heterogeneity in the prediction. The bimodal distribution (primary mode centered at 0.65) encompasses 60 samples and a secondary concentration of 50 samples around negative 0.15. It suggests the presence of two distinct application behavior categories. The positive correlation cluster that represents over 75% of samples indicates successful temporal pattern extraction through the model’s 48-integer input window, which corresponds to a 4-hour historical context given the 5 -minute integer granularity. This window size is sufficient to capture short-term dependencies while maintaining computational efficiency. The negative correlation subset that comprises 20% of samples represent applications with aperiodic usage patterns that challenge the model’s assumption of temporal regularity inherent in the architecture.

[0141] Overall, results demonstrated that adopting examples of a progressive decomposition for vehicular cache prediction may achieve improvements in temporal pattern recognition and cache efficiency. The consistent performance across popularity quartiles and robust validation trajectories confirm that example architectures may successfully translate time series forecasting into practical cache optimization.

[0142] From the foregoing it will be appreciated that, although specific embodiments have been described herein for purposes of illustration, various modifications may be made while remaining with the scope of the claimed technology.

[0143] Examples described herein may refer to various components as ‘"coupled” or signals as being “provided to” or “received from” certain components. It is to be understood that in some examples the components are directly coupled one to another, while in other examples the components are coupled with intervening components disposed between them.Similarly, signal may be provided directly to and / or received directly from the recited components without intervening components, but also may be provided to and / or received from the certain components through intervening components.

Claims

Docket No. 0077145-07501CLAIMSWhat is claimed is:

1. A system comprising:cache memory in a vehicle;at least one processor in the vehicle; andat least one computer readable media encoded with executable instructions which, when executed by the at least one processor, cause the sy stem to perform operations comprising:form a personal area network comprising the vehicle and one or more user devices, the personal area network formed using a wireless communication protocol;forecast certain data to be cached into the cache memory, based on activity in the personal area network; andcache the certain data over the personal area network from the one or more user devices into the cache memory of the vehicle.

2. The system of claim 1, the operations further comprising:select the certain data from the personal area network to be cached into the cache memory, wherein the operation select the certain data comprises selecting most recently used data (mru), selecting least recently used data (Iru), selecting earliest-deadlme first data (erf), selecting latest-deadline first data (Idf), selecting high-processing first data (HPF), or combinations thereof.

3. The system of claim 1, wherein the certain data originates from a source, the source comprising one or more of an access point, a cloud core, a mobile station, or the one or more user devices.

4. The system of claim 3, the operations further comprising:preload the certain data from the source into the vehicle prior to the operation form the personal area network.

5. The system of claim 1, wherein the operation forecast the certain data comprises:generate future demand data; andcache the future demand data over the personal area network from the one or more user devices into the cache memory' of the vehi cle.Docket No. 0077145-075016. The system of claim 1, wherein the activity' in the personal area network comprises application usage generated by the one or more user devices.

7. The sy stem of claim 5, wherein the operation generate the future demand data comprises:receive a request that is sent from the one or more user devices, wherein the request comprises a request for the certain data being used by least one user device of the one or more user devices at a time period;timestamp the request with the time period to create a timestamped request; send the timestamped request to the cache memory', wherein:if the timestamped request does not satisfy a condition set forth by a subset state of the cache memory,instruct generation of an age counter for each incremented subset state of a plurality of incremented subset states, wherein the each incremented subset state comprises the subset state incremented by an integer,instruct generation of a demand forecast parameter, wherein the demand forecast parameter comprises a predicted demand for a subsequent request at a subsequent time period; andcache the demand forecast parameter into the cache memory of the vehicle.

8. The system of claim 7, wherein the condition comprises a presence or an absence of the certain data requested for in the request,9. The system of claim 7, wherein the integer is five minutes,10. The system of claim 7, wherein the subsequent time period is one hour after the time period.

11. The system of claim 7, wherein the timestamped request is a first timestamped request, and the operations further comprising:receive a second timestamped request from the one or more user devices;insert the second timestamped request into the cache memory, andin response to the inserting the second timestamped request, evict a non-compliant timestamped request that does not satisfy an eviction parameter, wherein the eviction parameter comprises the demand forecast parameter and the age counter.

12. The system of any one of claims 1 to 11, wherein the personal area network is isolated to the vehicle, the one or more user devices, and the wireless communication protocol.Docket No. 0077145-0750113. The system of any one of claims 1 to 11, wherein the personal area network is isolated to the vehicle and the one or more user devices.

14. The sy stem of any one of claims 1 to 11, the operations further comprising:authenticate the one or more user devices to generate an authenticated one or more user devices, wherein the operation authenticate comprises:input a command for the one or more user devices to provide an authentication parameter to the personal area network; andupon a successful authentication, wherein the successful authentication comprises the one or more user devices satisfying the authentication parameter, access the personal area network with the authenticated one or more user devices.

15. The system of claim 14, wherein the personal area network is isolated from the one or more user devices that have had an unsuccessful authentication, wherein the unsuccessful authentication comprises the one or more user devices failing the authentication parameter.

16. The system of any one of claims 1 to 11, further comprising at least one antenna coupled to the vehicle, wherein the at least one antenna is configured for wireless communication with the one or more user devices to form the personal area network.

17. The system of claim 16, wherein the at least one antenna communicates with the one or more user devices at a bandwidth between 6-60 GHz.

18. The system of any one of claims 1 to 11, further comprising an array, the array comprising one or more antennae, wherein the one or more antennae communicates with the one or more user devices at a bandwidth between 6-60 GHz.

19. The system of any one of claims 3 to 11, wherein the personal area network is isolated from the cloud core.

20. The system of claim 19, wherein the cloud core comprises one or more of an edge computing server, cloud server, or fog computing server.

21. The system of any one of claims 3 to 11, wherein the personal area network is isolated from the access point.

22. The system of any one of claims 3 to 11, wherein the personal area network is isolated from the mobile station.Docket No. 0077145-0750123. The system of any one of claims 1 to 11, wherein the vehicle includes a power source configured to power the at least one processor.

24. The sy stem of any one of claims 1 to 11, the operations further comprising:receive a future request for the certain data that is sent from the one or more other user devices; andservice the future request from the cache memory.

25. A method comprising:forming, using a computing device comprising a memory and at least one processor, a personal area network including a vehicle and one or more other user devices, the personal area network formed using a wireless communication protocol;forecasting, using the computing device, certain data from the personal area network to be cached into a cache memory of the vehicle; andcaching, using the computing device, the certain data over the personal area network from the one or more user devices into a cache memory of the vehicle.

26. The method of claim 25, wherein the forecasting comprises:generating, using the computing device, future demand data, wherein the future demand data comprises a predicted demand for a future request at a subsequent time period, wherein the subsequent time period occurs later than an initial time period.

27. The method of claim 26, wherein the generating the future demand data comprises: receiving, using the computing device, a request for the certain data that is sent from one or more user devices, the request comprising a timestamp;sending, using the computing device, the request to the cache memory; and generating, using the computing device, a plurality of counters for a corresponding plurality of future subset states of a cache memory, wherein each future subset state of the corresponding plurality of future subset states comprises an initial subset state of the cache memory incremented by an integer.

28. The method of any one of claims 25 to 27, further comprising:isolating, using the computing device, the personal area network to the vehicle and the one or more user devices, the isolating comprises:severing, using the computing device, a connection between the personal area network and a source, the source comprising one or more of an access point, a cloud core, a mobile station, or one or more unauthorized user devices.Docket No. 0077145-0750129. The method of claim 28, wherein prior to the isolating the personal area network to the vehicle and the one or more user devices, further comprising offloading, using the computing device, the future demand data onto the vehicle.30, A computer-implemented method of training a machine learning model to generate a forecast model, the method comprising:receiving, using a computing device comprising a memory and at least one processor, a plurality of request vectors, each request vector of the plurality of request vectors comprising a request vector for an application at a time period;decomposing, using the computing device, the plurality of request vectors, to generate a plurality of trend vectors, and a plurality of seasonal vectors;passing, using the computing device, the plurality of seasonal vectors through a selfattention block, wherein the self-attention block identifies correlation scores, producing a plurality of forecast vectors;mapping, using the computing device, the plurality of forecast vectors to a plurality of predicted demand vectors, andgenerating, using the computing device, the forecast model by comparing the plurality of predicted demand vectors to a plurality of ground- truth aggregated demand vectors,31, The method of claim 30, wherein the recei ving comprises receiving the plurality of request vectors from a user device connected to a personal area network.

32. The method of claim 30, further comprising:applying, using the computing device, the forecast model to a cache memory, wherein the cache memory assigns a priority rating to a plurality of future request vectors received by the computing device, based on the plurality of request vectors,33. The method of any one of claims 30 to 32, wherein the plurality of request vectors comprises data from one or more of the following: most recently used data (mru), least recently used data (Iru), earliest-deadline first data (erf), latest-deadline first data (Idf), or high-processing first data (HPF),34. The method of any one of claims 30 to 32, wherein the comparing the plurality of predicted demand vectors to the plurality of ground-truth aggregated demand vectors comprises minimizing an error between the plurality of predicted demand vectors and the plurality of ground-truth aggregated demand vectors over a horizon.

Images