text
stringlengths
0
89.3k
corporation of gradientbased learning in SbPGs reduces training
times and achieves more optimal policies than its baseline
Index Terms Gradientbased optimization distributed learn
ing statebased potential games smart manufacturing systems
machine learning
I I NTRODUCTION
The emergence of the Internet of Things IoT cyber
physical systems and artificial intelligence AI has trans
formed production environments which enhances efficiency
and adaptability 1 These technologies have revolution
ized various sectors by improving productivity quality and
safety while reducing costs For instance IoT sensors enable
realtime monitoring of production processes 2 Similarly
AIpowered algorithms boost decisionmaking as in 3 In
manufacturing production systems are often characterized
by complex control parameters and considered multiagent
systems MAS 4 with multiobjective optimization chal
lenges 5 Such complexity requires agile and adaptive control
mechanisms which can be realized through the adoption of
distributed architectures with decentralizing control systems
Distributed production systems are applicable to a wide range
of sectors within the process industry eg food production
chemical plants pharmaceuticals and the transportation of
grains and coal
Recent years have witnessed a huge demand for self
learning capabilities within distributed production systems 67 through learning from experiences Advanced methodolo
gies have been developed to meet this demand Deep learn
ing particularly through multiagent reinforcement learning
RL 8 is the foremost among these For instance multi
agent RL for energyoptimal production policies 9 RLbased
scheduler for dynamic manufacturing jobs 10 and federated
RL for distributed energy management 11 Despite such
advancements realworld applications remain limited due to
lengthy training times and complex training processes involv
ing numerous parameters and settings 12 Furthermore the
effectiveness of multiagent RL in collaborative environments
is arguable as agents often operate independently and focus
on optimising individual objectives Consequently a challenge
in MAS is in enabling collaborative behaviour among agents
aligned with overarching system objectives
An alternative selflearning approach gaining traction in
distributed learning MAS involves integrating game theoretical
GT methods as in 1315 Among these statebased
potential games SbPGs with best response learning 15
have appeared particularly successful which offer a simpler
structure than deep learning and better suitability for real
world applications Despite their success there is room for im
provement notably in optimizing the exploration mechanism
during policy training to faster convergence while improving
performance In best response learning 15 players operate
random sampling during the exploration phase which may
lead to inefficiencies and lack of exploration direction which
increases training times and slows policy convergence
Hence our study aims to introduce a novel method for en
hancing SbPGs by integrating gradientbased learning which
effectively guides players exploration direction leading to
faster and smoother convergence compared to best response
learning We also propose solutions to deal with the chal
lenge of unknown and nonconvex utility functions inherent
in gradientbased learning by developing multiple estimation
function variants Here are the contributions of our study
We introduce gradientbased learning within SbPGs for
selflearning in distributed production systems which
employs Gradient Ascent with Newtons First Divided
Difference Method
We present three variants of estimation function to acarXiv240610015v1 csLG 14 Jun 2024commodate various characteristics of the objective func
tion which are the basic method augmented with mo
mentum and incorporating polynomial interpolation
We validate the proposed methods through application in
a laboratory testbed and a comparative analysis with best
response learning
The paper is structured as follows In Sec II we provide
preliminary research Sec III presents the fundamentals of
SbPGs and best response learning Sec IV elaborates on
the proposed gradientbased learning within SbPGs Sec V
presents the results and discussions of our experimentation
Finally in Sec VI we draw conclusions based on our findings
II P RELIMINARY RESEARCH
In this section we explore relevant literature including GT
for distributed optimization and gradientbased optimization
A Game Theory for Distributed Optimization
In recent years GT has expanded beyond its origin domain
in economics to contain engineering disciplines 16 which
plays a crucial role in analyzing and optimizing complex
systems involving multiple decisionmakers such as resource
allocation 17 autonomous driving 18 and cyber security
protocols 19 GT has facilitated selflearning within dis
tributed production systems as shown in SbPGs 15 Previous
studies 20 21 have spread foundational frameworks for
analyzing the convergence of equilibrium points in SbPGs
and defining conditions for dynamic and adaptive games
which have further extended for distributed selflearning MAS
in 15 Moreover we have expanded SbPGs by integrating
communication and memorybased learning mechanisms 22
as well as modelbased learning domains 6 23 However
a limitation remains due to the dependency on best response
learning 15 where random actions are chosen uniformly
during exploration This results in slower and unstable learn
ing behaviours Hence our study aims to enhance training
efficiency and smoothness through gradientbased learning