text
stringlengths
0
89.3k
while preserving SbPGs as foundational game structures of
distributed selflearning
B Gradientbased Optimization
Gradient ascent 24 a cornerstone of optimization involves
iteratively adjusting parameters to maximize an objective
function by following its positive gradient direction with
variants 25 eg stochastic and batch gradient ascent In
contrast gradient descent 24 operates oppositely Various
optimization algorithms like Adam RMSprop and Adagrad
provide strategies for efficiently updating model parameters
thereby improving convergence in optimization tasks 26
These also find application in machine learning which serves
as the backbone for training neural networks 26 opti
mizes support vector machines 27 and fits regression mod
els 28 However when applied to nonconvex problems chal
lenges arise in locating global optima among multiple local
minimamaxima 24 Overcoming these challenges requires
strategies like warm restarts 29 and annealing schedules 30Moreover gradientbased optimization has been theoretically
proven to achieve faster and smoother convergence 31 as
demonstrated in applications such as convolutional neural
networks 32 and nonlinear optimization methods 33 to
name a few In this study we present gradientbased optimiza
tion techniques within SbPGs to achieve faster convergence
improve the balance between exploration and exploitation and
enhance resource utilization compared to the undirected nature
of random sampling
III F UNDAMENTALS OF STATE BASED POTENTIAL GAMES
AND BESTRESPONSE LEARNING
In this section we explain the fundamental game structure
of SbPGs and its learning algorithm best response learning
A Statebased Potential Games
We start with the fundamental principles of SbPGs followed
by their application in optimizing distributed manufactur
ing processes SbPGs represent a subset of potential games
PGs 34 A PG contains a set of Nplayers denoted individu
ally as i 1 N Unlike other games where each players
utility function uiis influenced solely by their individual
actions ai in PGs the utility function uiin PGs is influenced
by the global state of the entire system An additional com
ponent of PGs is a scalar potential function ϕ also known as
the global objective function Consequently the strategicform
game of PGs can be formulated as ΓN Aui ϕ where
Arepresents the set of individual player actions ai
Furthermore in SbPGs 15 the set of states Sand the state
transition process Pare integrated into the game To ensure
SbPGs hold to the fundamental theorem of PGs the potential
function ϕaS Rcimust be available within the game
structure Furthermore each actionstate pairing a sAS
must satisfy the following conditions
uiai suia
i ai s ϕai sϕa
i ai s1
and
ϕai sϕai s 2
for any state sinPa sas well as Rcidenotes continuous
actions
In 15 we demonstrate their application in manufacturing
optimization settings and prove the existence of its convergent
as illustrated in Fig 1 The distributed system under consider
ation is characterised and examined using graph theory 35
Further details regarding SbPGs in a production chain of
manufacturing graph are presented in 6 15 22 The
overarching goal is to optimize production systems in a fully
distributed manner avoiding reliance on a centralized instance
which remains consistent throughout this study
B Best Response Learning
In 15 best response learning was developed as an algo
rithm for training the policy of each player within SbPGs A
performance map is used to record the players experiences
and visualize the distribution of actions across the state space
as shown in Fig 1 This map discretizes the state space intoFig 1 An overview of selflearning mechanism in distributed
production systems using a game structure of SbPGs
support vectors l 1 L Each support vector retains the
bestexplored and its utility value within the related state The
performance maps are updated as follows
ul
imaxsi ul
iksi iful
iksi ul
imaxsi 3
al
imaxsi al
iksi iful
iksi ul
imaxsi 4
in which ul
ikandal
ikrepresents the utility and action resulted
by player iin the current iteration k associated with the lth
support vector
During exploration actions al
iksiare computed based on
random uniform sampling which results in utilities ul
iksi
This process allows agents to gather experiences and store
the bestexplored actions in the maps Meanwhile during
exploitation each player iselects an action ait1by globally
interpolating the performance map according to its current
state The global interpolation rule is managed as follows
DS0Sk