Stefan Bosse - Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

Chapter 13

Energy Management

Low-power Design and Smart Energy Management

Power Analysis and Algorithmic Selection 470

Smart Energy Management with Artificial Intelligence 480

Further Reading 488

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

470

This chapter introduces different techniques for the optimization and the

low-power design of application-specific data processing and agent process-

ing platforms. Furthermore, agent-based energy management is introduced,

in addition to the energy management SoMAS introduced in Chapter 9.

13.1 Power Analysis and Algorithmic Selection

In contrast to various other energy management approaches targeting

algorithms and architectures with high computational effort, Smart Energy

Management (SEM) can be performed spatially at run-time by applying a

dynamic selection from a set of different (implemented) algorithms classified

by their demand of computational power, and temporally by varying data pro-

cessing rates. The smart energy management can be implemented with

decision trees, based on Quality-of-Service (QoS) and energy constraints. It

can be shown that the power and energy consumption of an application-spe-

cific SoC design strongly depends on the computational complexity of the

used algorithms.

For example, a classical Proportional-Integral-Differential (PID) controller

used for the feedback position control of an actuator requires basically only

the P-part; the I- and D-parts only increase position accuracy and response

dynamics, which are selectable. Depending on the actual state of the system

and the actual and estimated future energy deposit, suitable algorithms can

be selected and executed optimizing the QoS and the trade-off between accu-

racy and economy.

Signal and control processing can be modelled on abstract algorithmic level

using signal flow diagrams. These signal flow graphs are mapped to Petri Nets

to enable direct high-level synthesis of digital SoC circuits using a multiprocess

architecture with the Communicating-Sequential-Process model on execution

level and the high-level synthesis framework ConPro.

Power analysis using simulation techniques on digital gate-level provides

input for the algorithmic selection at run-time of the system leading to a

closed-loop design flow. Additionally, the signal-flow approach enables power

management by varying the signal flow rate, which will be discussed later.

The smart energy management using algorithmic selection relates to the

entire design flow for low-power SoC data processing and control systems,

introduced in Chapter 12, and which should be demonstrated in the following

using a concrete example. The system is modelled on an abstract level using

signal flow diagrams [BOS10D].

Figure 13.1 shows a composition of a complete feedback-controlled system

consisting of sensor signal acquisition (ADC), filtering, an error controller with

a proportional, integral, and differential sub-controller [ISE89], and finally a

signal generator (DAC) driving an actuator.

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.1 Power Analysis and Algorithmic Selection 471

Fig. 13.1 Composition and modelling of a digital control system with signal flow dia-

gram [BOS11B]

The controller is used to control the position of an actuator. The control

error is defined by the difference of the acquired position signal X(S) and the

desired position parameter S’.

This initial specification (see also Section 5.5) is used to derive 1. A multi-

process programming model; and 2. A hardware model for an SoC design on

for the energy optimization at synthesis- and run-time.

The signal flow diagram is first transformed into an S/T Petri Net rep-

resentation, which is shown in Figure 13.2. Functional blocks are mapped to

transitions, and states represent data, which is exchanged between those

functional blocks. The partitioning of functional blocks to transitions of the net

can be performed at different composition and complexity levels. The signal

flow diagram from Figure 13.1 was partitioned using complex blocks (com-

posed of low-level blocks like multipliers and adders) to reduce

communication complexity (and data processing latency).

Sensor data (X) is acquired periodically and passed to the data processing

system. A token of the net is equal to a data set of one computation pro-

cessed by the functional blocks. The functional blocks P, I, and D are placed in

concurrent paths of the net.

ADC DACPID

-1

1 -

-1

(1- )K

-1

USX

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

472

Fig. 13.2 Mapping of the signal flow diagram to a Petri Net and mapping of Petri Net to

communication channels and sequential processes using the ConPro pro-

gramming language [BOS11B].

The Petri Net is then used 1. To derive the communication architecture, and

2. To determine an initial configuration for the communication network. Func-

tional blocks with a feedback path require the injection of initial tokens in the

appropriate states (not required in the example).

States of the net are mapped to buffered communication channels and

transitions are mapped on concurrently executing processes - each with

sequential instruction processing - using the ConPro programming language,

shown in Figure 13.2, too. Details are discussed in Section 5.5.

Energy Analysis

The derived multiprocess programming model was synthesized to a digital

logic SoC using high-level synthesis. For simulation, gate-level synthesis was

performed with a standard logic cell technology library. The resulting net-list

was analysed with an event-driven simulator, calculating the overall cell activ-

ity for each time unit, defined by terms of cell output changes. Synthesis and

analysis were performed using the Concurrent Programming (ConPro) com-

piler and the Silicium-Compiler-and-Analyser framework (SiCA).

P I D

channel c_x:int[12] with model=buffered;

channel c_s:int[12] with model=buffered;

channel c_e1,c_e2,c_e3:int[12] ...

channel c_u1,c_u2,c_u3:int[12] ...

process p_i:

begin

reg t1,t2,z1,z2: int[DATAWIDTH4];

always do

begin

t1 <- c_e2;

t1 <- z1, z1 <- t1;

t1 <- t1 * KI;

t1 <- t1 asr 4;

t2 <- t1 + z2,

z2 <- t2;

c_u2 <- t2;

end;

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.1 Power Analysis and Algorithmic Selection 473

Fig. 13.3 Averaged SoC cell activity correlates strongly with computation and signal/

data flow. After obtaining the fifth result value U, the I and D computational

blocks are enabled. [BOS11B]

The SoC circuit activity correlates strongly with a computation of a new data

set (with sensor data input sampled periodically) and the computation com-

plexity, shown in Figure 13.3.

The logic cell activity of the circuit has strong peaks around the computa-

tion of a new output value U. About every 140 clock cycles a new input value X

is generated, triggering the calculation of a new output value U.

The first five data sets are computed with an enabled P-part of the control-

ler only. After the fifth computation, the I and D parts were enabled, too. This

results in an increase of circuit activity of about 50%.

Unfortunately, the power dissipation cannot be estimated directly from this

cell activity. Logic cells consist of a network of (paired) transistors. Power dissi-

pation of a CMOS circuit depends proportionally from the transistor switching

activity. Simulation results for the controller are shown in Figure 13.4 (using

SiCA, too). There is only weak correlation between data processing activity

(and computation complexity) and power dissipation due to clocking activity

of registers.

Power dissipation can only be estimated from the above circuit cell activity

if clock-gated registers are assumed [XIA02]. The principle-architecture of

gated registers is shown in Figure 13.5. The clock gating prevents switching

activity (and hence power dissipation) inside the register cell in the case of no

change of input data (D=Q).

0 200 400 600 800 1000

Clock Cycles

100

120

140

Cell Activity

Clock Cycle

Digcon1

Processing of

a data set with

simplified algorithm

Processing of

a data set with

different algorithm

+50% activity

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

474

Fig. 13.4 Averaged SoC transistor switching activity of the circuit retrieved from simula-

tion. Power dissipation is proportional to transistor activity. [BOS11B]

Fine-grained gate-level clock gating is a requirement for the proposed low-

power design method and enables again a strong correlation between com-

putation activity (and hence algorithmic complexity) with the power

dissipation of the data processing system.

Results of such a modified control system with clock-gated registers are

shown in Figure 13.6. There is again a significant increase of transistor switch-

ing activity of about 30% if the two different computation levels (P, PID) are

compared.

The demonstrated correlation of the algorithmic and computational com-

plexity with the energy and power consumption assumed an optimized ASIC

design technology process.

Fig. 13.5 Register Clock Gating (adapted from [XIA02]).

200

400

600

800

1000

Clock Cycles

200

400

600

800

1000

1200

1400

Transistor Activity

Clock Cycle

Digcon1

Conventional

CLK

Clock-Gated

CLK

≠

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.1 Power Analysis and Algorithmic Selection 475

Fig. 13.6 Averaged SoC transistor switching activity of the circuit retrieved from simula-

tion. Using clock-gated registers results again in strong correlation between

data processing activity (and computation complexity) and power dissipation.

[BOS11B]

In [TAS13] the algorithmic correlation was investigated for a SRAM-LUT-

based FPGA technology, showing again a significant algorithmic correlation

suitable for run-time energy management based on algorithmic selection.

Smart Energy Management at Run-time

Self-powered systems must deal with a limited amount of energy during

run-time. The energy charge in future is uncertain. Smart energy manage-

ment should handle the conjunction of energy demand and energy

conversation.

Methods from Artificial Intelligence (AI) can be used to manage energy at

runtime with dynamic parameter adaptation and algorithmic selection. AI

methods differ in complexity, thus only few are suitable to be embedded in

microchips. Suitable methods are for example constraints nets and decision

trees in conjunction with machine learning approaches.

A simulation of a complex sensor-actuator system implementing the PID

controller from Figure 13.1 should demonstrate the benefits of using a deci-

sion tree method for dynamic parameter adaptation, which can be retrieved

from machine learning. Parameters to be controlled are data processing rate

R and the algorithmic level L (1: only P, 2: P+I+D controller). The controller per-

forms minimization of the position error of the actuator, that means the

difference between a desired and a measured (angular) position.

200

400

600

800

1000

Clock Cycles

100

200

300

400

500

Transistor Activity

Clock Cycle

Digcon1

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

476

Fig. 13.7 Decision tree used for energy optimization at runtime based on parameters.

Input parameters: Error, Quality=Run-time/Error. Output parameters: Data

processing rate R={1,5,10,20,100}, Algorithmic level L={P:1,PID:2}

The system is charged with a stochastic energy source and discharged by

computation and actuator activity. The power and energy required for the

computation can be calculated from the results of the power analysis, the

power and energy required for actuator activity can be calculated by a simpli-

fied physical simulation.

The decision tree used in the simulated system is shown in Figure 13.7. A

decision tree is system- and environment-specific and must be derived for

each different system.

Parameters are modelled with a discrete set of values related to a discrete

set of cost values, shown in Definition 13.1.The cost values are used to calcu-

late the overall runtime costs of the system. The goal is to minimize the

overall costs and to maximize the energy conversation, but still serving the

quality of the service to be provided (in this case the accurate position control

of the actuator).

Def. 13.1 System parameter value sets and related cost values.

DataprocessingrateR={1,5,10,50,100}[milliseconds]

R_C={100,50,10,5,1}

AlgorithmiclevelL={1,2}

L_C={100,150}derivedfrompoweranalysis

ActuatorpositionerrorE={0,5,10,100}[arb.units]

E_C={0,250,500,5000}

The smart energy algorithms used for the system simulation are shown in

Algorithm 13.1. The estimate procedure calculates actual computation and

quality-of-service costs for the system, and the choose procedure calculates

optimized values for the data processing rate and algorithmic level based on

QualityQuality

L=2

R=5

Error

L=2

R=10

L=2

R=10

L=2

R=20

L=1

R=100

<=5

>=0.0003<0.0003

>= 0.0003<0.0003

>30

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.1 Power Analysis and Algorithmic Selection 477

actual system state. The costs are derived from the previous power analysis

results.

The costs function returns a linear interpolated cost value of the particular

parameter.

The basic concepts of the simulation are shown in Algorithm 13.2. The

charge procedure stores energy in the system from a random source. The con-

troller procedure implements the selectable P/PI(D) controller, and finally the

stimuli procedure simulates the simplified mechanical actor behaviour due to

a drive signal calculated by the controller.

Alg. 13.1 Smart energy management algorithm based on algorithmic selection

[BOS11B]

VAR

level,rate:actualalgorithmiclevelanddataprocessingrate

error:actualpositioncontrolerror

runtime:passedruntimeintimeunits

energy:energystorageinarb.units

cost:qualityofservicecosts

PROCEDURE

Estimate:

foreachtimeunitdo

delta:=Costs(level)*Costs(rate)+Costs(error);

energy:=energy‐delta;

quality:=Average(runtime)/Average(cost);

runtime:=runtime+1;

cost:=cost+Costs(error);

Choose(level,rate);

PROCEDURE

Choose:

Usedecisiontreetochooseoptimal{level,rate}valuesbasedon

averagedqualityandactualerror

Alg. 13.2 System simulation algorithms [BOS11B]

VAR

pos_act,pos_set:actualanddesiredactuatorposition

PROCEDURE

Stimuli:

each500thtimeunitdo:pos_set:=‐pos_set;

pos_act:=pos_act+drive/4;

delta:=pos_act‐pos_set;

error:=min100(absdelta);

PROCEDURE

Charge:

energy:=energy+Random(CHARGE_MAX)

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

478

PROCEDUREController:

caselis

when1:P‐controller

foreachratetimeunitdo:

delta:=pos_act‐pos_set;

S:=delta*KD/100;

drive:=S;

when2:PIcontroller

foreachratetimeunitdo:

delta:=pos_act‐pos_set;

S:=delta*KD/100;

Z1’:=Z1,Z2’:=Z2,Z1:=delta;



T:=(Z1’*KI)/100+Z2’;

Z2:=T;

drive:=S+T;

Figure 13.8 shows simulation results of the complex sensor-actuator system

implementing the PID controller. The system runs always out of energy if a

fixed parameter setting {Rate, Level} is used, regardless of the parameter val-

ues. In contrast the system can reach a stable state balancing energy charging

and discharging if dynamic parameter adaptation based on the decision tree

method and system feedback is used.

Fig. 13.8 System simulation with different runtime behaviours using a decision tree,

which can be retrieved by machine learning methods. Parameters: Data pro-

cessing rate={1,5,10,20,100}, Algorithmic level={P:1,PID:2} [BOS11B]

500

1000

1500

2000

200 000

400 000

600 000

800 000

1. x 10

Energy [Arb. Units]

Time [Arb. Units]

Dynamic Run Using

Decision Tree Parameter

Optimization

Static Parameter Run

Rate=20, Level=1

Rate=10, Level=1

Rate=20, Level=2

Rate=100, Level=1

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.1 Power Analysis and Algorithmic Selection 479

13.1.1 Low-power Agent Processing

Dynamic algorithmic selection can be used to optimise the power consump-

tion of the agent processing at run-time. The power consumption for each

activity of an ATG of a specific agent class AC can be analysed. This analysis

marks all activities with a power weight, which can be used at run-time to esti-

mate the energy required for execution of this specific activity. Agents are

capable to adapt their behaviour based on these power weights and energy

information provided by the host, i.e., by reading an energy and energy man-

agement tuple from the his tuple-space, which was stored by a stationary

energy management agent.

Application-Specific PCSP Platform

In the case of the hardware implementation of the non-programmable

PCSP Agent Processing Platform the power analysis is performed with gate-

level simulation techniques, which provides input for the calculation of the

power weights. Usually the simulation of the entire application specific APP is

too complex. Therefore, each activity is decomposed in elementary state-

ments, which can be simulated separately.

In the case of the software implementation, the machine code is analysed

and profiled for some representative test patterns.

The Agent-based Platform simulation techniques can be used to estimate

the power weights of activities roughly for some representative test patterns.

This method gives the best balance between power analysis accuracy and

analysis complexity, which can be significantly large in the case of the hard-

ware simulation.

Programmable PAVM Platform

The programmable platform approach cannot be optimized in this way with

respect to a specific algorithm due to its generic nature. But with the analysis

of the energy requirement of specific machine instructions it is possible to

optimise the overall VM design with a back propagation of the energy infor-

mation to the software AFM/AML compiler energy optimized programs can be

created (design time). Furthermore, at run-time the energy demand can be

estimated based on the energy profiling, enabling the power efficient schedul-

ing of different agents. An agent token can be associated with an energy

parameter that can be used to provide input for decision making process

regarding the execution of an agent.

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

480

13.2 Smart Energy Management with Artificial Intelligence

13.2.1 Introduction

Sensorial materials equipped with embedded miniaturized smart sensors

provide environmental information required for advanced machine and

robotics applications. With increasing miniaturization and sensor-actuator

density, decentralized self-supplied energy concepts and energy distribution

architectures are preferred and required.

Self-powered sensor nodes collect energy from local sources, but can be

supplied additionally by external energy sources. Nodes in a sensor network

can use communication links to transfer energy, for example, optical links are

capable of transferring energy using Laser or LE diodes in conjunction with

photo diodes on the destination side, with a data signal modulated on an

energy supply signal.

A decentralized sensor network architecture is assumed with nodes sup-

plied by 1. energy collected from a local source, and 2. by energy collected

from neighbour nodes using smart energy management (SEM). Nodes are

arranged in a two-dimensional grid with connections to their four direct

neighbours. Each node can store collected energy and distribute energy to

neighbour nodes.

Each autonomous node provides communication, data processing, and

energy management. There is a focus on single System-On-Chip (SoC) design

satisfying low-power and high miniaturization requirements.

Energy management is performed 1. for the control of local energy con-

sumption, and 2. for collection and distribution of energy by using the data

links to transfer energy.

Typically, energy management is performed by a central controller in that a

program is implemented [LAG10], with limited fault robustness and the

requirement of a well-known environment world model for energy sources,

sinks, and storage. Energy management in a network involves the transfer of

energy.

The loss of energy  (in the range between 0 and 1) at each node occurring

each time when “energy” is routed along different nodes from a source to a

destination node (assuming N intermediate nodes) reduces overall efficiency

dramatically in the order of =

By using electrical connections, only negligible loss of energy can be

expected in a distributed network, in contrast to optical and radio wave con-

nections, which have significant loss in the order of  10-30% per node.

Additionally, in the latter case there is no physical interaction between a

source and a sink node requesting energy, thus requiring active management

(routing).

To overcome these limitations and to increase operational robustness,

smart energy management is performed by using concepts from artificial

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.2 Smart Energy Management with Artificial Intelligence 481

intelligence. Initially, the sensor network is a distributed group of independent

computing nodes. Interaction between nodes is required to manage and dis-

tribute information and energy. One common interaction model is the mobile

agent.

Different kinds of agents with different behaviours are used to negotiate

energy demands and energy distribution and to implement group communi-

cation. A multi-agent system is a decentralized and self-organizing approach

for data processing in a distributed system like a sensor network.

Recent work shows the benefit and suitability of multi-agent systems used

for energy management [LAG10].

13.2.2 Communication, Data and Energy Transport

Network nodes, arranged for example in a two-dimensional mesh-network

(see Figure 13.9), can exchange data with their neighbours by using serial

communication links. There are different kinds of physical transmission tech-

nologies available: electrical, optical, and radio-wave based.

In contrast to electrical interconnect technologies, optical and radio-wave

technologies have the disadvantage of lower efficiency . This is negligible for

the exchange of information, but significant for the distribution and exchange

of energy required for the supply of nodes. Optical communication has clear

advantages such as extremely small and light-weight hardware, ultra-low

power consumption, and the ability to optimally focus and match the beam to

the transmission medium (optical fibre) [KED06].

Sharing of one interconnect medium for both data communication and

energy transfer significantly reduces node and network resources and com-

plexity, a prerequisite for a high degree of miniaturization required in high-

density sensor networks embedded in sensorial materials. Point-to-point con-

nections and mesh-network topologies are preferred in high-density

networks because they allow good scalability (and maximal path length) in the

order of (log N), with N as the number of nodes.

Figure 13.9 shows the main building blocks of a sensor node, the proposed

technical implementation of the optical serial interconnect modules, and the

local energy management module collecting energy from a local source, for

example a thermo-electric generator, and energy retrieved from the optical

communication receiver modules. The data processing system can use the

communication unit to transfer data (D) and superposed energy (E) pulses

using a light-emitting or laser diode. The diode current, driven by a differen-

tial-output sum amplifier, and the pulse duration time determine the amount

of energy to be transferred. The data pulses have a fixed intensity several

orders lower than the adjustable energy pulses. On the receiver side, the

incoming light is converted into an electrical current by using a photo diode.

The data part is separated by a high-pass filter, the electrical energy is stored

by the harvester module.

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

482

Fig. 13.9 Network topology (left) and sender and receiver blocks (right) used for data

and energy transmission between neighbour nodes that are connected in the

network.

Information and energy is carried with agents that are encapsulated in mes-

sages routed in the network from a source to a destination node using, for

example, a simple delta-routing path protocol.

13.2.3 Multi-Agent Interaction Model and Implementation

Having the technical abilities explained in the previous section, it is possible

to use active messaging to transfer energy from good nodes having enough

energy towards bad nodes, requiring energy. An agent can be sent by a bad

node to explore and exploit the near neighbourhood. The agent examines

sensor nodes during path travel or passing a region of interest (perception)

and decides to send agents holding additional energy back to the original

requesting node (action). Additionally, a sensor node is represented by a node

agent, too. The node and the energy management agents must negotiate the

energy request.

The agents have a data state consisting of data variables and the control

state, and a reasoning engine, implementing behaviours and actions. Table

Data Processing

Storage

Communication

Data

Storage

Sensor

Acquisition

Sensor Node

Harvester

Energy

Processing

Units

Sender Unit

Receiver Unit

Energy Harvester

Storage

Data Processing

Light Guide

Data Processing

Light Guide

Energy Harvester

Storage

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.2 Smart Energy Management with Artificial Intelligence 483

13.1 explains different agent behaviour required for the smart energy man-

agement (see page 485).

A node having an energy level below a threshold can send a help agent with

a delta-distance vector specifying the region of interest in a randomly chosen

direction. The help agent hops from one sensor node to the next until the

actual delta-vector is zero. If there is a good node found, with local energy

above a specified threshold, the help agent persists on this node and tries to

send periodically deliver agents transferring additional energy to the origina-

tor node. An additional behaviour, help-on-way, changes the deliver agent

into an exploration agent, too. Such a modified agent examines the energy

level of nodes along the path to the destination. If a bad node was found, the

energy carried with the agent is delivered to this node, instead to the original

destination node. A more general AAPL agent model of this SoS is discussed in

detail Section 9.3.

This approach is independent of the agent processing platform and the

mobile agents can be processed on both proposed platforms (PAVM, PCSP).

Originally, this SEM approach was implemented on simplified application-spe-

cific processing platforms, shown in Figure 13.10.

Fig. 13.10 Sensor node building blocks providing mobility and processing of multi-agent

systems: Parallel agent virtual machines, agent-processing scheduler, commu-

nication, and data processing.

AGENT

STATE

AGENT VM AGENT VM AGENT VM

A-Queue A-Queue A-Queue

COMMUNICATION

MESSAGE

POOL

SCHEDULER

AGENT

STATE

AGENT

STATE

DATA PROC.

M-QueueM-Queue

Send

Receive

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

484

Fig. 13.11 UML activity diagram for the Smart Energy Management Agent behaviour.

In this early approach the state of an agent was completely kept in the mes-

sage transferred in the network, but not the functional behaviour. Figure

13.10 shows the execution environment for the energy management agents.

There is a message module implementing delta-distance routing (see Section

4.3.2 for a discussion), and several finite-state-machines implementing the

agent behaviour that provide virtual machines capable to process incoming

DIR

DX*

DY*

ENERGY E

AGENT TYPE T

AGENT STATE S

AGE

Interact Incoming

N:=MyNode()

E:=E*ε

N.E:=N.E+E

Percept/Request

S:=I AM HERE

Percept/Reply

S:=I AM HERE

Percept/Deliver

S:=I AM HERE

N.FLOW:=

-DirOfPos(DX,DY)::

N.FLOW

Percept/Help

S:=I AM HERE

DX*:=DX*-DX

DY*:=DY*-DY

Percept/Help

S:=DYING

AGE:=0

Percept/Dsitribute

S:=I AM HERE

Action/Reply

S:=DYING

AGE:=0

Action/Help

S:=DYING

Action/Request

AGE:=AGE-1

Action/Request

S:=DYING

AGE:=0

SEND(REPLY,

DX=-DX*,DY=-DY*

E=DE)

S:=DYING

Action/Deliver

S:=DYING

AGE:=0

AWAIT

Action/Help

AGE:=AGE-AGEBAD

Action/Help

AGE:=AGE-AGEGOOD

SEND(DELIVER,

DX=-DX*,DY=-DY*

E=DE)

Action/Distribute

S:=DYING

AGE:=0

ROUTEDISCARD

PERCEPT

ACTION

otherwise

S=DYING S=TRAVEL

otherwise

N.E+DEPOSIT>DE

otherwise

AGE<=0

T=HELP

AGE=0 |

N.QoS < THRES

otherwise

S=I AM HERE

T=DISTRIBUTE

N.E-DE>THRES

T=DELIVER

T=REPLY

T=REQUEST

N.E<THRES

DX=0 & DY=0

T=DISTRIBUTE

DX=0 & DY=0 &

N.E+DEPOSIT<DE

N.E+DEPOSIT>DE

T=HELP

HELPONWAY & N.E<THRES

DX=0 & DY=0

DX=0 & DY=0DX=0 & DY=0

T=DELIVER

T=REPLY

T=REQUEST

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.2 Smart Energy Management with Artificial Intelligence 485

agents. All parts are mappable on digital logic with RTL and SoC system

architectures.

Figure 13.11 shows the UML activity diagram of the reasoning engine, imple-

menting all types of agents, forming a super agent. The delta-distance or

region of interest vector (DX, DY) is modified on each message routing, the

(DX*, DY*) vector holds an unmodified copy, which is used for replies, the DE

entry specifies the requested energy, and the ENERGY entry reflects the actual

energy carried with the message (without the contribution of the data part

itself). This entry is altered each time a message is routed, respecting the

transmission efficiency .

13.2.4 Analysis and Experimental Results of the SEM

A first proof of concept and experimental results were achieved by using a

multi-agent simulation framework (SeSAM, see Section 11.1). The simulation

test-bed consists of a network with N=100 nodes arranged in a 10 by 10

Agent Type Behaviour

Request

Point-to-point agent: this agent requests energy from a specific

destination node, returned with a Reply agent.

Point-to-point agent: Reply agent created by a Request agent,

which has reached its destination node. This agent carries

energy from one node to another.

Help

ROI agent: this agent explores a path starting with an initial

direction and searches a good node having enough energy to

satisfy the energy request from a bad node. This agent resides

on the final good node for a couple of times and creates multi-

ple deliver agents periodically in dependence of the energy

state of the current node.

Deliver

Path agent: this agent carries energy from a good node to a bad

node (response to Help agent). Depending on selected sub-

behaviour (HELPONWAY), this agent can supply bad nodes first,

found on the back path to the original requesting node.

Distribute

ROI agent: this agent carries energy from and is instantiated on

a good node and explores a path starting with an initial direc-

tion and searches a bad node supplying it with the energy.

Tab. 13.1 Agents with different behaviour used to manage and distribute energy (ROI:

region of interest).

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

486

matrix. Each node can communicate with up to four neighbours in the direc-

tions DIR= {North, South, West, East}.

Each node N

periodically collects (store) energy from a local source having

a stochastically distributed energy spectrum in the range [0,E

i,max

]. Monte

Carlo simulation is used to specify each E

i,max

before a simulation run. Data

processing, interaction with the environment (e.g. sensor acquisition), and

agents consume energy, which reduces the energy deposit E

Each sensor node is modelled with an agent, too. Energy management

agents and sensor node agents negotiate energy demands and communicate

by using globally shared variables. If the energy deposit of a node is below a

threshold E

low

(called bad node), help agents are sent out, if E

high

good

then distribute agents are used to distribute energy to surrounding bad

nodes. If E

good

low

then the node is fully functional (called good node).

Assuming a specific stochastic spatial configuration {E

i,max

}, simulation

results in Figures 13.12 and 13.13 show the benefit of energy management

using help- and distribute agents. Without energy management, there are

about 50% bad nodes (blue rectangles) never reaching an energy level above

a critical threshold. With agents, the spatial energy distribution is more regu-

lar, and the fraction of bad nodes is below 10% all the time. The Figure 13.14

compares the number of bad nodes resulting from different combinations of

agents and agent behaviours.

Fig. 13.12 Simulation results for a network of sensor nodes arranged in 10 by 10 matrix

topology without any distributed energy management. Shown are the spatial

energy distribution after 10000 time steps (right) and the temporal population

for bad and good nodes (left).

E<250

E>500

E>250

E>1000

E>1500

E>2000

E>2500

E>3000

E>3500

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

13.2 Smart Energy Management with Artificial Intelligence 487

Fig. 13.13 (Left) Simulation results with energy management. Help, deliver, and distrib-

ute agents are used to compensate low-energy nodes. (Right) Agent

deployment in network

Using additional distribute agents results in a decrease of 30% relative to

the case using only help agents, but absolutely the benefit is below 5% and is

therefore negligible. Moreover, the fraction of all nodes with mean up-time

below a critical threshold (10%) is always below 5%.

The Figure 13.14 shows the temporal progress of total system energy in

dependency on different energy management agents, too. Due to the high

loss of energy transfer between nodes (here 20%), the total energy efficiency

is dramatically decreased compared with the case without management, and

distribute agents reduce the total system efficiency again about 50%.

The multi-agent implementation offers a distributed management service

rather than a centralized approach commonly used. The simple agent behav-

iours can be easily implemented in digital logic hardware (application-specific

platform approach).

To summarize, help agents with simple exploration and exploitation behav-

iours are suitable to meet the goal of a regular energy distribution and a

significant reduction of bad nodes unable to contribute sensor information,

but additional distribute agents create no significant benefit.

E<250

E>500

E>250

E>1000

E>1500

E>2000

E>2500

E>3000

E>3500

Help

Agent

Distribute

Agent

Deliver

Agent

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)

Chapter 13. Energy Management

488

Fig. 13.14 Comparison of total sensor network energy (left) and bad node population

(right) for different energy management agent systems

13.2.5 Algorithmic Selection and SEM

In Section 9.4 a self-organizing MAS was introduced that is capable to dis-

tribute energy between nodes of a sensor network. The algorithmic analysis

and selection approach introduced in the previous section can be a prerequi-

site for the specification of the various parameters of the mobile smart energy

management (SEM) and node energy management (SEN) agents. The SEM and

SEN agents make assumption about the cost of computation. The methods

from the previous section can be used to select different SEM and SEN behav-

iour and to compute the reward for energy distribution and help requests.

The algorithmic selection approach can be embedded in the agent transition

process to add energy saving goals that can be related to the original activity

and mobility goals.

13.3 Further Reading

1. D. Kedar and S. Arnon, Optical wireless communication in distributed sen-

sor networks, SPIE Newsroom, 2006.

2. S. Bosse, D. Lehmhus, W. Lang, M. Busse (Ed.), Material-Integrated Intel-

ligent Systems: Technology and Applications, Wiley, ISBN: 978-3-527-

33606-7 (2018)

3. Clemens Moser, Power Management in Energy Harvesting Embedded Sys-

tems, PhD Thesis, ETH Zürich, 2009, Shaker, ISBN 978-3832282059

4. S. Borlase, Ed., Smart Grids - Infrastructure, Technology, and Solutions,

CRC Press, 2013, ISBN 9781439829103

S. Bosse, Unified Distributed Sensor and Environmental Information Processing with Multi-Agent Systems

epubli, ISBN 9783746752228 (2018)