You can approach this assignment in many different ways.  We will not be giving any guidance on what algorithms are most appropriate.

Your solution must include a Probabilistic Graphical Model as the core component . This core

component should be your own implementation and can include any tutorial code, but it cannot simply be a sklearn implementation.  You can use any algorithm as part of your approach, including any algorithm available in Python’s sklearn library.

It is recommended you start this assignment by brainstorming several different possible approaches. Make sure you understand what information is available, what information is uncertain, and what assumptions it may be reasonable to make.

Every area on the floor plan is named with a string: r, c, or outside. The strings r and c stand for room and corridor, respectively.

Data

The files data1 .csv and data2 .csv contain 2 different complete data that is representative of a typical weekday in the office building. This data includes the output of each sensor as well as the true number of people in each room. This data was generated using a simulation of the building, and your program will be tested against many days of data generated by the same simulation. Because this data would be expensive to collect, you are only given two sets of 2400 complete data points, from two workdays. The simulation attempts to be a realistic approximation, so it includes many different types of noise and bias. You should treat this project as if the data came from a real office building, and is to be tested on real data from that building. You can make any assumptions that you think would be reasonable in the real world, and you should describe all assumptions in the report. Part of your mark will be determined by the feasibility of your assumptions, if applied to the real world.

The number of people who come to the office each day varies according to this distribution: num_people  =  round(Normal(mean=20,  stddev=3)) .

This information was obtained from records of the number of workers present each day, and the empirical

distribution of num_people was found to be identical to the above distribution.

Data format specification

Sensor data

Your submission file must contain a function called get_action(sensor_data), which receives sensor data in the following format:

sensor_data  =  { 'Motion_Sensor1 ' :  'motion ' ,  'Motion_Sensor2 ' :  'motion ' ,

'Motion_Sensor3 ' :  'motion ' ,  'Motion_Sensor4 ' :  'motion ' ,  'Motion_Sensor5 ' :  'motion ' ,                'Motion_Sensor6 ' :  'motion ' ,  'Motion_Sensor7 ' :  'motion ' ,  'Motion_Sensor8 ' :  'motion ' ,                'Motion_Sensor9 ' :  'motion ' ,  'Motion_Sensor10 ' :  'motion ' ,  'door_sensor1 ' :  0,                                'door_sensor2 ' :  0,  'door_sensor3 ' :  0,  'door_sensor4 ' :0,  'door_sensor5 ' :0,  'door_sensor6 ' :0, 'door_sensor7 ' :0,  'door_sensor8 ' :0,  'door_sensor9 ' :0,  'door_sensor10 ' :0,  'door_sensor11 ' :0,

'Camera1 ' :  0,  'Camera2 ' :  0,  'Camera3 ' :  0,  'Camera4 ' :  0,

'robot1 ' :  ( 'r1 ' ,  0),  'robot2 ' :  ( 'r8 ' ,  0),  'time ' :  datetime .time(8,  0)}

We have four different kinds of sensors for this building:

• The first one is the motion sensor. The output of this sensor is two states: motion, and no motion. motion means the sensor captured a motion in the room, and no motion means the sensor did not capture anything. We know that the performance of these sensors is not very accurate, but these sensors are cheap and accessible.

• The second sensor is the door sensor. These sensors count the number of people who go in or out the room (either way). So, the output of this sensor is a positive integer. Same as motion sensors, door sensors are not 100% accurate too.

• The third sensor is the camera sensor. These sensors are more accurate than the other ones and more expensive. As before, no sensor in the real world is completely accurate. The output of this sensor is the absolute number of people that it counts. Since these sensors are more expensive, very few rooms have them.

• We also have two robots in the building. The robots wander around the building and count the number of people in each room. The value is a 2-tuple of the current room, and the number of people counted. For example, if the robot goes into r4 and counts 8 people, it would have the value (‘r4’,8). If it goes into corridor‘c2’and no one is present, it would have value (‘c2’,0).

The value of time is a datetime.time object representing the current time. Datapoints will be provided in 15 second resolution, i.e., your function will be fed data points from 15 second intervals from 8 am - 6 pm.

Training data

Both data1 .csv and data2 .csv have the same structure.  They contain a column for each of the above sensors, as well as columns for each room, which tell you the current number of people in that room. The columns of data files are the following and can be divided into two groups:

1. Columns that represent readings from sensors, as described in the previous section..

2. Columns that are present only in the training data and provide the ground truth with the number of people in each room, corridor, and outside the building: r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, c1, c2, c3, outside. This ground truth data provides the instantaneous count of people per room (i.e. every

15 seconds, a snapshot of each room is magically taken at exactly the same time, and the number of people in each room is counted. If someone passes through multiple rooms within 15 seconds, they will not increment the count in multiple rooms, only in one room).

Note that the first column of data files is the index, and has no name.

You should use this data to learn the parameters of your model. Also, you can save the parameters to csv files that can be loaded during testing.

Action data

get_action() must return a dictionary with the following format. Note that every numbered room named r in the building has lights that you can turn on or off.  All other rooms/corridors have lights that are permanently on, which you have no control over, and which do not affect the cost.

actions_dict  =  { 'lights1 ' :  'off ' ,  'lights2 ' :  'off ' ,  'lights3 ' :  'off ' ,

'lights4 ' :  'off ' ,  'lights5 ' :  'off ' ,  'lights6 ' :  'off ' ,  'lights7 ' :  'off ' ,

'lights8 ' :  'off ' ,  'lights9 ' :  'off ' ,  'lights10 ' :  'off '}

The outcome space of all actions is (on,off).

In the provided example_solution .py, there is an example code stub that shows an example of how to set up your code.

Figure 1 shows the floor plan specification.  Please notice a door between room 3 and the outside area (entrance door), but there is no sensor in this door.

Figure 1: Floor plan.

Cost specification

If a light is on in a room for 15 seconds, it usually costs you 1 cent. If there are people in a room and there is no light on, it costs you 4 cents per person every 15 seconds, because of lost productivity. The cost can be calculated exactly using the complete training data, so it is also based on an instantaneous count of the number of people in each room.

Your goal is to minimise the total cost of lighting plus lost productivity, added up over the whole day. You do not need to calculate this cost, the testing code will calculate it using the actions returned by your function,

and the true locations of people (unavailable to you). The file example_test .py shows exactly how the cost is calculated.

Testing specification

Your program must be submitted as a python file called solution .py. During testing, solution .py will be placed in a folder with test .py. A simpler version of test .py has been provided (called example_test .py), so you can confirm that testing will work. A more elaborate version of test .py will be used to grade your solution.

There is a strict upper time limit for the final submission.  If your submission runs for longer than 1800 seconds for 10 days (180s/day), it will be cut off, and you will receive 0 marks for the programming section of the assignment.

The file solution .py will be reloaded for each new day (using importlib.reload), so you may do any daily setup in that file outside of the get_action function, and it will still work.

We will run a test evaluation one week before the deadline, and release the cost and time of each students

model. This is to help you confirm that your model does not have any major errors, and works with the evaluation system.  To participate, make a submission for the assignment before 1 week before the final deadline. We will make an announcement to remind you. To make sure you have something to submit, try to make sure you have at least a minimum working model by this time.