Hamed Asadi thinks machines can learn a lot from humans.
The UA graduate research assistant’s work in cognitive radio engines, or CEs, often draws parallels between an infant’s growth into adulthood and an engine’s artificial intelligence: going so far as to label the four progressive learning stages of a CE “infant, childhood, teenager and adult.”
“In the infant stage, the engine has a lot to learn and is at its most unstable,” Asadi said. “By the time it reaches the adult stage, the engine has learned most of what it needs to maintain a steady link and perform optimally.”
Asadi’s training algorithm even incorporates what he calls the “forgetfulness factor,” allowing the engine to delete or forget older, less useful information about possible configurations. The less knowledge it has to cycle through, the faster the engine will remember and adjust its settings. All this adds up to a small device -- like a mobile phone -- that seamlessly detects and switches to available wireless connections (no more manual switching from Wi-Fi to data).
For the second year in a row, Asadi’s research earned him an award at WInnComm, an annual conference hosted by Wireless Innovation Forum on wireless communication and software-defined or cognitive radio technologies. Last year Asadi’s paper was one of four to receive Best Paper. This year he alone won Paper of the Year.
The paper details Asadi’s training algorithm, which he says will stabilize a cognitive engine’s performance during its learning stages and reduce the chances for dropped signals or lost data.
Cognitive Engines Grow Up
In years past an operator manually adjusted a radio or antenna -- settings such as the antenna strength, frequency or power consumption -- for a stronger signal or faster connection. Cognitive engines have mostly automated the process of learning and selecting the best radio settings.
A cognitive engine’s learning phase is like a series of tests by trial and error. The engine selects a setting or single configuration and repeatedly sends data until it reaches a point where it no longer transmits effectively. Then it tries another configuration, and so on, until it finds a combination that works the best for its prevailing needs. Simultaneously the engine monitors the surrounding external network environment, checking for available signals and possible obstructions, settings that can’t be altered so easily.
Just as with humans, learning takes time and energy, and even the best multitaskers can get overwhelmed.
“Wireless environments especially are very dynamic. The simplest thing, like a building or too many people trying to access the same network, can interfere with the signal and disrupt transmission,” said Asadi. “If the radio is operating in a fast-changing environment it may not have the luxury of time to learn what is best, significantly impacting its performance and connection.”
One example of a radio trying to adjust too quickly, and failing, is a streaming video in a moving car. Chances are spotty video or constant buffering will result from an antenna seeking, receiving then losing a signal.
“We have the metrics. We have the algorithm. Now we are working on applying them in training scenarios to improve cognitive engines’ performance,” said Asadi. “Every improvement means more reliability for next-generation technologies.”
Hamed Asadi’s research is supported by the Broadband Wireless Access & Applications Center based at the UA.