Wireless Innovation Forum Top Ten Most Wanted Innovations
Innovation #4: Artificial Intelligence/Machine Learning for Radios
4.1 Executive Summary
With the advent of artificial intelligence and machine learning technologies being broadly applied to many aspects of life, it’s only natural for the technology to be applied to radios, as well. There are a number of potential use cases, commercial, civilian and military, some of which are already under investigation. Examples include:
- The DARPA Spectrum Collaboration Challenge which focuses on the use of machine learning for radios to maximize interference mitigation and spectral usage.
- Research to enable radios to change or adapt the physical layer in order to increase the effectiveness of transmission while substantially reducing resource utilization and power consumption.
- Research to enable proactive dynamic spectrum management in order to optimize spectrum utilization and foresee spectrum demand.
- Programs so that radios can autonomously identify signals of interest, including potential threats.
In order for these scenarios to come to fruition, there are a number of enabling technologies and ecosystems that need to be incubated, nurtured and developed. There is more work needed to be done with regards to both software and hardware. On the software side, one of the greatest challenges is to implement the learning databases, with on-going updates, and then algorithms to make the best decisions using the learning databases. The development of the actual Artificial Intelligence/Machine Learning (AI/ML) algorithms and datasets themselves, form the core of the cognition ability of the radio. On the hardware side, power efficiency, cost per unit, ease for programming and memory requirements all come into play and must be weighted to find the most optimal solution. Different types of processors can be viable choices depending on the specific use cases.
4.2 Application
Developing and deploying autonomous radios capable of AI/ML will benefit a variety of use cases for commercial, civilian and military purposes.
- AI/ML will drive systems for automation and network evolution. Network operators’ Operating Expenses (OPEX) can be drastically reduced and customer experience greatly improved as what matters the most gets efficiently delivered.
- AI can be applied to the military, first responders and public safety communities to augment analysis and decision-making capabilities and reaction times both, speeding up learning, and improving their ability to act with discretion, accuracy, and care under uncertain and changing conditions.
- Regulators will have the potential to enable a close to real time and proportionate regulatory regime that identifies and addresses risk while also facilitating far more efficient regulatory compliance.
4.3 Description
AI/ML for radios will be enabled by both hardware and software innovations. For hardware, there is a need for embedded platforms which can support multiple AI/ML use cases and requirements. In terms of processors, the primary choices today for AI/ML processing are Application Specific Integrated Chips (ASIC), (Graphics processing Unit) GPU and Field Programmable Gate Arrays (FPGA). In general, ASIC is the lowest in power efficiency and cost per unit, GPUs are high performance but consume a lot of power, and FPGAs are harder to program but consume less power than GPUs. In addition, memory requirements tend to be high for AI/ML systems in order to store the datasets for training, as well as cognition. And there are also multiple potential form factors desired by different end users, including small standalone devices to rack-mounted units more rugged form factors like VPX. Since there is no single architecture that can support all the disparate requirements of each potential use case, there is a clear opportunity for multiple vendors to compete in AI/ML COTS hardware platforms.
With regards to software, one of the greatest challenges is to implement the learning databases, with on-going updates, and then algorithms to make the best decisions using the learning databases. The development of the actual AI/ML algorithms and datasets themselves, form the core of the cognition ability of the radio. Other software key needs include development related to programming tools and methodologies, libraries and frameworks. Examples include:
- Tools for programming FPGAs and GPUs which make them easier to program for AI/ML algorithms. There are tools today which support programming languages like OpenCL, which are often used by users starting with GPUs as the key processor. Even FPGAs now support OpenCL and C programming methodologies in order to be easier to program and compete against GPUs.
- Libraries and frameworks to leverage for quicker development of AI/ML systems. A good example of this is Google’s TensorFlow.
- Then there are the actual AI/ML algorithms and datasets themselves, which form the core of the cognition ability of the radio.
There are plenty of opportunities in the ecosystem for any of these types of innovations that make it easier to develop and deploy radios capable of AI/ML. As the enabling technologies develop and the ecosystem matures, we will see the advent of fully autonomous radios that support a variety of use cases for commercial, civilian and military purposes.