Explainable AI Whiteboard Technical Series: Reinforcement Learning

0:09 radio frequency environments are

0:11 inherently complex and therefore

0:13 challenging to control and optimize for

0:15 the efficient transmission of data since

0:18 the Inception of radio frequency or RF

0:20 radio Resource Management also known as

0:23 RRM has been a long-standing technique

0:25 used to optimize the RF radio waves that

0:28 transmit Network traffic in Wireless

0:30 plans however multiple interference

0:32 sources like walls buildings and people

0:35 combined with the air servings of

0:36 transmission medium make RRM a

0:38 challenging technique to

0:40 master traditionally site surveys have

0:43 been used to determine the optimal

0:44 placement of Wi-Fi access points and

0:47 settings for transmit power channels and

0:49 bandwidth however these manual

0:51 approaches can account for the dynamic

0:53 nature of the environment when the

0:55 wireless network is in use with people

0:57 and devices entering or leaving and

0:58 moving about additionally this challenge

1:01 is compounded with random RF

1:03 interference from sources like microwave

1:05 ovens radios and aircraft radar to name

1:07 a

1:08 few but what if the wireless network

1:11 itself could perform RRM on its own what

1:14 if it could detect and respond to both

1:16 interference sources as well as the

1:18 movement of people and devices and

1:20 adjust the radio settings in real time

1:21 to provide the best possible wireless

1:23 service that's exactly what Juniper's

1:26 done with the AI native misted Wireless

1:28 solution using it Advanced machine

1:30 learning techniques specifically misuses

1:32 reinforcement learning to perform

1:35 RRM in a nutshell a reinforcement

1:37 learning machine or agent learns through

1:40 an iterative trial and error process in

1:42 an effort to achieve the correct results

1:44 it's rewarded for actions that lead to

1:46 the correct result while receiving

1:48 penalties for actions leading to an

1:50 incorrect result the machine learns by

1:52 favoring actions that result in rewards

1:55 with missed Wireless the reinforcement

1:57 learning machine's value function is

1:59 based B on three main factors that lead

2:01 to a good user experience coverage

2:04 capacity and connectivity a value

2:06 function can be thought of as an

2:08 expected return based on the actions

2:09 taken the machine can execute five

2:12 different actions to optimize the value

2:14 function these are adjusting the band

2:16 settings between the two wireless bands

2:18 of 2.4 GHz and 5 GHz increasing or

2:22 decreasing the transmit power of the

2:23 ap's radios switching to a different

2:26 Channel within the band adjusting a

2:28 Channel's bandwidth and switching the

2:30 BSS color which is a new knob available

2:32 to 11 ax access points RRM will select

2:36 actions with maximum future rewards for

2:38 a site future rewards are evaluated by a

2:42 value function the various actions taken

2:44 by the Learning machine such as the

2:46 increase of transmit power or switching

2:48 the band from 2.4 gigs to 5 gigs

2:51 together represent a policy which is a

2:53 map the machine builds based on multiple

2:55 trial and error Cycles as it collects

2:57 rewards modeling actions that maximize

2:59 the value function function again keep

3:01 in mind that the value function

3:03 represents good Wireless user experience

3:06 as time goes on even if random changes

3:08 occur in the environment the machine

3:10 learns as it strives to maximize the

3:12 value function the benefits of using

3:14 reinforcement learning are obvious a

3:17 missed wireless network customizes the

3:19 RRM policy per site creating a unique

3:22 wireless coverage environment akin to a

3:24 well-tailored

3:26 suit while large organizations with

3:28 multiple sites replicate their many

3:30 locations as copy exact these sites will

3:33 naturally experience variances despite

3:36 best efforts reinforcement learning

3:38 easily fixes this delivering realtime

3:41 actively adjusting custom Wireless

3:43 environments we hope this episode help

3:46 to uncover some of the magic and mystery

3:48 behind our AI native Network Solutions