Juniper Mist Location Services

0:00 foreign

0:06 SDK is responsible for receiving the ble

0:10 RSSI information and combining it with

0:13 sensor data from the mobile device and

0:16 sending that back to the cloud every

0:18 second while the mobile device is moving

0:21 in this machine Learning System

0:22 continuously adapts to changing RF

0:25 environments without the need for site

0:27 surveys

0:29 unlike single antenna ble designs Mist

0:32 uses a single transmitter to drive eight

0:35 unique directional antennas each beam

0:38 contributes to the likely location of

0:40 the device the Miss technology examines

0:42 all the probability surfaces one for

0:45 each directional beam and combines them

0:47 to find the most likely point in the map

0:50 where the device is located

0:52 so two terms of importance are ple and

0:55 intercept

0:56 ple is the path loss exponent Mist uses

1:00 this formula to determine the expected

1:02 signal strength at various locations on

1:04 a map based on a gain at that direction

1:07 from the antenna

1:09 taking into account orientation and

1:12 ceiling height

1:13 The Intercept is another constant that

1:15 must be derived it is much like The

1:18 Intercept of a line and indicates the

1:20 expected power at one meter from the

1:22 antenna

1:24 individual RF environments and even

1:27 device types have different optimal ple

1:30 and intercept values this ensures the RF

1:33 model adapts to the RF environment as it

1:36 changes and accounts for the differences

1:37 between mobile device types providing a

1:40 consistent user experience

1:42 the essential element behind miss

1:44 machine learning is to seek the maximum

1:47 agreement between the results of

1:49 increasingly many location estimates

1:51 from that construct individualized path

1:55 loss formulas

1:56 as the machine learning algorithm begins

1:59 it's using the default plf for a device

2:02 type meaning an iPad an iPhone 6 an

2:05 iPhone 6s

2:07 machine learning is running continuously

2:10 the only time you would see an AP at

2:13 zero learning is when it was just added

2:15 to the map or something has been changed

2:18 in its configuration regarding ceiling

2:20 height orientation Etc

2:23 the thing to look for here is to make

2:25 sure all of your APS are learning to

2:27 reach level one usually only takes a

2:30 couple minutes for a device and that's

2:32 the most important level because 95

2:34 percent of the work is done reaching

2:37 level one learning after that the

2:39 algorithm is looking for new unique data

2:42 points

2:43 we have created this visual to help you

2:45 quickly understand the status of

2:47 location machine learning

2:49 if an AP is not learned on a device it

2:52 will be gray when all the devices are

2:54 learned they will show Green

2:56 so what is going on here using the show

2:59 all button you can see the raw data and

3:01 note you can also choose to expose just

3:04 a single device

3:06 what the UI has done

3:08 is to add up the ple and intercept then

3:11 a medium value is derived per device and

3:13 per AP of variance from that median is

3:16 shown

3:17 in a normal environment we would expect

3:19 this variance not to be more than one or

3:22 two if it is three or four the tile will

3:25 show yellow and more than five it will

3:27 be red

3:28 this is to alert you to take a look at

3:30 why this AP may be so different from the

3:33 others

3:34 so in this case this is what I would

3:37 expect

3:38 if you saw a five or a six something

3:41 that's kind of radically different from

3:42 the others immediately I would go look

3:44 at the floor plan and see if that AP is

3:47 in a strange spot that would justify it

3:50 having a much different plf than the

3:52 rest of the aps

3:58 foreign