A Multiagent System Approach to Schedule Devices in Smart Homes - - PowerPoint PPT Presentation

A Multiagent System Approach to Schedule Devices in Smart Homes

Ferdinando Fioretto University of Michigan

Feb, 2017

William Yeoh Enrico Pontelli New University State University

1 Fig.1 Fig.2

Home Automation

2 Fig.3

Network of smart homes

Outline

• Background (DCOPs)
• Smart Homes Device Scheduling (SHDS)
• Results
• Conclusions and Future work

3

Background SHDS Results Conclusions

Distributed Constraint Optimization

4

Constraint graph fab fbc fac xc xb xa Constraint (cost table)

<X, D, F, A, α>:

• X: Set of variables.
• D: Set of finite domains for each variable.
• F: Set of constraints between variables.
• A: Set of agents, controlling the variables in X.
• α: Mapping of variables to agents.

xa xb cost 3 1 ∞ 1 2 1 1 5 Background SHDS Results Conclusions

Distributed Constraint Optimization

5

<X, D, F, A, α>:

• X: Set of variables.
• D: Set of finite domains for each variable.
• F: Set of constraints between variables.
• A: Set of agents, controlling the variables in X.
• α: Mapping of variables to agents.
• GOAL: Find a cost minimal assignment.

x⇤ = arg max

x

F(x) = arg max

x

X

f2F

f(x|scope(f))

F

min min

Background SHDS Results Conclusions

DCOP: Assumptions

• Agents coordinate an assignment for

their variables.

• Agents operate distributedly.

Communication:

• By exchanging messages.
• Restricted to agent’s local neighbors.

Knowledge:

• Restricted to agent’s sub-problem.
• Privacy preserving.

6

fab fbc fac xc xb xa xd fbd

Background SHDS Results Conclusions

Smart Home Device Scheduling (SHDS)

A SHDS problem is composed of:

• H: A neighborhood of smart homes.
• Zi : A set of smart electric devices within each home hi.
• H: A time horizon for the device scheduling.

7 exact 2 Zi. The ener

problem uildings hi and whose

Background SHDS Results Conclusions

Smart Home Device Scheduling (SHDS)

A SHDS problem is composed of:

• H: A neighborhood of smart homes.
• Zi : A set of smart electric devices within each home hi.
• H: A time horizon for the device scheduling.
• θ : A pricing function expressing cost per kWh of energy consumed.

8

Background SHDS Results Conclusions

Smart Home

A smart home hi has:

• A set of smart devices Zi it can control.

9

exact 2 Zi. The ener

problem uildings hi and whose

Background SHDS Results Conclusions

Smart Home

A smart home hi has:

• A set of smart devices Zi it can control.
• Li A set of locations (e.g., living room, kitchen)

10

Background SHDS Results Conclusions

Smart Home

A smart home hi has:

• A set of smart devices Zi it can control.
• Li A set of locations (e.g., living room, kitchen)
• PH A set of state properties (e.g., cleanliness, temperature)

11

Thermostat Cleanliness sensor Battery charge sensor Background SHDS Results Conclusions

Smart Devices (Actuators)

A smart device is defined with a

• Location, defining the place where the device can act (e.g., living room)
• The possible actions it can perform (clean, charge, stop) and the power

consumption associated to them

• The set of state properties it affects (e.g., cleanliness, battery_charge)

12 Action State property Power (kW/h)

run cleanliness 0.0 charge battery charge 0.26 stop 0.0

Background SHDS Results Conclusions

Smart Devices (Sensors)

• We associate a predictive model to each home sensor.
• It describes the transition of a state property from a state s and

time t to time t+1, when affected by a set of actuators.

13

Thermostat

Heater Oven Current State Next State

• ff
• ff

12 C 11 C

• ff

bake 12 C 13.8 C

• n
• ff

12 C 17.5 C

• n

bake 12 C 19.3 C

Effect of the environment Background SHDS Results Conclusions

Smart Device Schedules

Scheduling Rules

• Simple syntax to express scheduling rules:
• Active rules: specify user-defined objectives on a desired state
• f the home. E.g.,
• Passive rules: define implicit constraints on devices. E.g.,

14

hlocationi hstate propertyi hrelationi hstatei htimei

living room cleanliness ≥ 75 before 1800

zv battery charge ≥ 0 always zv battery charge ≤ 100 always

Background SHDS Results Conclusions

Smart Device Schedules

Schedule: A sequence of actions for the home devices.

15

1400 1500 1600 1700 1800 15 30 45 60 75

Cleanliness (%)

15 30 45 60 75

Battery Charge (%) Time Goal Deadline

40 15 R 15 30 R 35 30 C 55 30 C 30 45 R 5 60 R 25 60 C 75 R 65 S

Device Schedule Cleanliness (%) Battery Charge (%)

Background SHDS Results Conclusions

min

ξ[0!H]

Zi

↵c·Csum + ↵e·Ediff

Smart Home Device Scheduling (SBDS)

• SHDS objective:

16

Aggregated monetary cost of the homes schedules Energy consumption peaks across all homes subject to: Homes’ devices schedules

8hi 2 H, R[ta!tb]

p

2 Ri : ⇠[ta!tb]

Φp

| = R[ta!tb]

p

P

All scheduling rules must be satisfied Background SHDS Results Conclusions

DCOP mapping

SBDS

• A home hi ϵ H.
• A device zj (in building hi)
• Action j for device zj.
• Schedule costs for a device zj
• Device scheduling feasibility
• Energy peak consumption

DCOP

• Agent ai ϵ A
• Variable xi ϵ X (controlled by ai)
• j-th value in domain Di of

variable xi

• Local soft constraint
• Local hard constraint
• Global soft constraint

17

Background SHDS Results Conclusions

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

1. Agents independently search for a feasible schedule for their local devices.

18 exact 2 Zi. The ener

problem uildings hi and whose ci : schedule cost Eit :energy consumption

Background SHDS Results Conclusions

exact 2 Zi. The ener

problem uildings hi and whose

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

1. Agents independently search for a feasible schedule for their local devices. 2. Schedule costs and energy consumption are broadcasted to all other agents.

19

(ci , Eit ) (ci , Eit ) ci : schedule cost Eit :energy consumption

Background SHDS Results Conclusions

exact 2 Zi. The ener

problem uildings hi and whose

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

3. Upon receiving all other agents costs and energy consumptions:

20

ci : schedule cost Eit :energy consumption (cj , Ejt ) (ck , Ekt )

Background SHDS Results Conclusions

exact 2 Zi. The ener

problem uildings hi and whose

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

3. Upon receiving all other agents costs and energy consumptions:

• Computes the objective cost with its current schedule.

21

ci : schedule cost Eit :energy consumption

lem with its current solution, αc · ci(ξ[0!H]

Zi

) + αe · Ediff Then, within a given time limit,

current schedule

Background SHDS Results Conclusions

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

3. Upon receiving all other agents costs and energy consumptions:

• Computes the objective cost with its current schedule.
• Within a time limit, it finds a new solution to its local subproblem that is no worse than

the current solution.

22

ci : schedule cost Eit :energy consumption

αc · ci(ˆ ξ[0!H]

Zi

) + αe · ˆ Ediff ≤ αc · ci(ξ[0!H]

Zi

) + αe · Ediff

current schedule new schedule

Background SHDS Results Conclusions

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

3. Upon receiving all other agents costs and energy consumptions:

• Computes the objective cost with its current schedule.
• Within a time limit, it finds a new solution to its local subproblem that is no worse than

the current solution.

• It computes the gain Gi between its current and new solutions, and broadcast it to all
• ther agents.

23

ci : schedule cost Eit :energy consumption

Gi = ⇣ αc · ci(ξ[0!H]

Zi

) + αe · Ediff⌘ − ⇣ αc · ci(ˆ ξ[0!H]

Zi

) + αe · ˆ Ediff⌘

Background SHDS Results Conclusions

exact 2 Zi. The ener

problem uildings hi and whose

Solution Approach

SH-MGM: Adaptation of a local search DCOP algorithm (MGM).

• 4. Upon receiving all other agents’ gains Gk, it checks if the agent has the largest

gain among all those received. If so, it updates its schedule to the new schedule, otherwise it retains its old schedule.

• 5. The process repeats untill convergence (all gains = 0) or a fixed number of

iterations.

24

Gi : agent’s gain Gj Gk

Background SHDS Results Conclusions

Evaluation: Settings

• 7 Raspberry Pis connected via a LAN.
• Each controlling 9 smart actuators.
• Communication and coordination of the MAS is implemented

via the JADE framework.

• Each agent uses an internal CP solver (JaCoP) to solve its local

scheduling problem.

25

A Raspberry PI with a dangle Smart devices Background SHDS Results Conclusions

Evaluation

SH-MGM vs. Uncoordinated approach.

26

Settings:

• H = 12 (step = 30 min)
• Realistic device consumptions and environment settings
• CP timeout = 10 sec

10 20 30 40 50 740 780 820 Number of Cycles Solution Cost

• SH−MGM

Uncoordinated

Main Results:

• SH-MGM finds better

solutions than a simple uncoordinated approach.

• Solution quality increases

with the number of cycles. Background SHDS Results Conclusions

Conclusions and Future Work

• We formalized the Smart Home Device Scheduling Problem

and cast it as a DCOP.

• We propose SH-MGM, a local search-based algorithm to solve

SHDS problems.

• Some results:
• SH-MGM finds better solutions than a simple uncoordinated method.
• Feasibility established for using a local search-based approach

implemented on hardware with limited storage and processing power.

• Future work:
• More realistic setting for the SHDS agents and devices.
• Taking account user preferences for the scheduling rules.

27

Background SHDS Results Conclusions

References:

• Fig. 1: http://goo.gl/5znqip
• Fig. 2: goo.gl/dqwUz2
• Fig. 3: goo.gl/WFzMhv

Thank You!