tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
as described in Fig 5 Each station features a silo and hopper
with maximum capacities of 1742L and 91L respectively
where the actuators facilitate the transfer of materials between
buffers Additionally a PLCbased Siemens ET200SP control
system is integrated into each station which facilitates com
munication via Profinet Furthermore a group of sensors is
installed to monitor the current level of each buffer which are
later used as states in SbPGs The BGLP includes a feature
that utilizes the You Only Look Once YOLO v8 model to
detect foreign objects in the system and activate an ejection
system when they are identified 40
In SbPGs each actuator is considered as one player The
utility function Uiformulated in 6 15 22 is designed to
combine the multiple objectives of the system which remains
in this study The utility function Uiis formulated as follows
Ui1
1 αlLip1i N1
1 αlLn
i1iN1
1αdVD1
1 αpPi
15
in which Piis the power consumption of actuator i1iN
denotes the identity function where iNshows the last
player in the sequence VDrepresents the fulfilled production
demand and αlαd and αpdenote weighting parameters for
each objective Constraints Li
sandLi
pare computed to prevent
overflow and bottleneck based on the upper and lower limits
of the corresponding buffers level To be noted the utility
function Uiis integral to the system see Fig 1 Therefore the
gradientbased learning player cannot access its utility function
and must resort to estimation methods outlined in Sec IV
In our experiments we implement a continuous production
process with a production demand set at 0110Ls Each
method undergoes a maximum of 20 training episodes and
1 testing episode with each episode lasting 10000 seconds
Each player computes a new action every 10 seconds Addi
tionally we discretize the state space into 40 Furthermore all
parameters related to the algorithms and learning approaches
pass automated tuning using Hyperopt 41 Then the simula
tion model of the BGLP is publicly accessible through both the
MLPro 42 and MLProMPPS 43 frameworks In this study
we conduct policy training within the simulation model due to
considerations of safety cost and time constraints However
the policy trained in the simulation can be deployed in the real
system within our laboratory
B Benchmark SbPGs with Best Response Learning
We apply the SbPGs approach with best response learn
ing 15 as our benchmark for comparison with the proposed
gradientbased learning method Through experimentation we
achieve the best results when training the players over 20TABLE I Results and comparisons between best response and
gradientbased learning in the BGLP
Method
EkoTraining
Time sOverflow
LsPower
kWsDemand
LsPotential

Benchmark Best Response Learning
200000 00000 04759 00000 129052
Gradientbased Learning
1 θk 0 200000 00000 04374 00000 124245
1 θk 3 120000 00000 04317 00000 124275
2 θk 0 200000 00000 04415 00000 130255
2 θk 4 120000 00000 04477 00000 124286
3 θk 0 180000 00000 04495 00000 124422
3 θk 1 110000 00000 04442 00000 123847
episodes In the testing episode we observe complete avoid
ance of overflow with an average power consumption of
0475885 kWs accomplishment of the production demand
of 0110Ls and an average potential value of 12905199
These outcomes suggest that the resulting policies derived
from SbPGs with best response learning are nearly optimal
However the training duration remains lengthy and the explo
ration behaviour remains uncontrollable
C Results on Gradientbased Learning in SbPGs
In this subsection we present the results of the gradient
based learning in SbPGs for all three variants considering
both with and without kickoff for each variant The number
of kickoff episodes is denoted as θk Table I summarizes
the testing results of all variants and provides comparisons
between the best response and gradientbased learning in the
BGLP
In the first variant without kickoff hyperparameter tuning
determined that the optimal parameter combination includes
αset to 10 and the OU noise ranging approximately within
03 Conversely with kickoff αremains at 10 OU noise
is deactivated and θkis set to 3 The testing results indicate a
reduction in power consumption by approximately 9 for both
approaches compared to the benchmark which achieves com
plete overflow avoidance and fulfilling production demand
Additionally kickoff episodes contribute to a 40 reduction
in training time compared to the benchmark
In the second variant both with and without kickoff
episodes involve deactivation of the OU noise with αset to
05 and βto 04 In the variant with kickoff θkis defined as
4 Test results demonstrate a similar trend to the first variant

as described in Fig 5 Each station features a silo and hopper

with maximum capacities of 1742L and 91L respectively

where the actuators facilitate the transfer of materials between

buffers Additionally a PLCbased Siemens ET200SP control

system is integrated into each station which facilitates com

munication via Profinet Furthermore a group of sensors is

installed to monitor the current level of each buffer which are

later used as states in SbPGs The BGLP includes a feature

that utilizes the You Only Look Once YOLO v8 model to

detect foreign objects in the system and activate an ejection

system when they are identified 40

In SbPGs each actuator is considered as one player The

utility function Uiformulated in 6 15 22 is designed to

combine the multiple objectives of the system which remains

in this study The utility function Uiis formulated as follows

Ui1

1 αlLip1i N1

1 αlLn

i1iN1

1αdVD1

1 αpPi

15

in which Piis the power consumption of actuator i1iN

denotes the identity function where iNshows the last

player in the sequence VDrepresents the fulfilled production

demand and αlαd and αpdenote weighting parameters for

each objective Constraints Li

sandLi

pare computed to prevent

overflow and bottleneck based on the upper and lower limits

of the corresponding buffers level To be noted the utility

function Uiis integral to the system see Fig 1 Therefore the

gradientbased learning player cannot access its utility function

and must resort to estimation methods outlined in Sec IV

In our experiments we implement a continuous production

process with a production demand set at 0110Ls Each

method undergoes a maximum of 20 training episodes and

1 testing episode with each episode lasting 10000 seconds

Each player computes a new action every 10 seconds Addi

tionally we discretize the state space into 40 Furthermore all

parameters related to the algorithms and learning approaches

pass automated tuning using Hyperopt 41 Then the simula

tion model of the BGLP is publicly accessible through both the

MLPro 42 and MLProMPPS 43 frameworks In this study

we conduct policy training within the simulation model due to

considerations of safety cost and time constraints However

the policy trained in the simulation can be deployed in the real

system within our laboratory

B Benchmark SbPGs with Best Response Learning

We apply the SbPGs approach with best response learn

ing 15 as our benchmark for comparison with the proposed

gradientbased learning method Through experimentation we

achieve the best results when training the players over 20TABLE I Results and comparisons between best response and

gradientbased learning in the BGLP

Method

EkoTraining

Time sOverflow

LsPower

kWsDemand

LsPotential

Benchmark Best Response Learning

200000 00000 04759 00000 129052

Gradientbased Learning

1 θk 0 200000 00000 04374 00000 124245

1 θk 3 120000 00000 04317 00000 124275

2 θk 0 200000 00000 04415 00000 130255

2 θk 4 120000 00000 04477 00000 124286

3 θk 0 180000 00000 04495 00000 124422

3 θk 1 110000 00000 04442 00000 123847

episodes In the testing episode we observe complete avoid

ance of overflow with an average power consumption of

0475885 kWs accomplishment of the production demand

of 0110Ls and an average potential value of 12905199

These outcomes suggest that the resulting policies derived

from SbPGs with best response learning are nearly optimal

However the training duration remains lengthy and the explo

ration behaviour remains uncontrollable

C Results on Gradientbased Learning in SbPGs

In this subsection we present the results of the gradient

based learning in SbPGs for all three variants considering

both with and without kickoff for each variant The number

of kickoff episodes is denoted as θk Table I summarizes

the testing results of all variants and provides comparisons

between the best response and gradientbased learning in the

BGLP

In the first variant without kickoff hyperparameter tuning

determined that the optimal parameter combination includes

αset to 10 and the OU noise ranging approximately within

03 Conversely with kickoff αremains at 10 OU noise

is deactivated and θkis set to 3 The testing results indicate a

reduction in power consumption by approximately 9 for both

approaches compared to the benchmark which achieves com

plete overflow avoidance and fulfilling production demand

Additionally kickoff episodes contribute to a 40 reduction

in training time compared to the benchmark

In the second variant both with and without kickoff

episodes involve deactivation of the OU noise with αset to

05 and βto 04 In the variant with kickoff θkis defined as

4 Test results demonstrate a similar trend to the first variant