text
stringlengths
0
89.3k
iq
S0xSkx2 S0ySky2 5
wS0Sk
i1
DS0Sk
i2γ 6
ait1X
kwS0Sk
iP
mwS0Sm
iak
i 7
in which xand ydenote the prior and subsequent states
nrepresents the player identifier and DS0Sk
i indicates the
distance from a support vector to the current state combination
while wS0Sk
i calculates the weighted value based on this
distance Here S0represents the current state combination Sk
denotes the state combination of the centre k and γrepresents
the smoothing parameter for the performance map
IV G RADIENT BASED LEARNING IN STATE BASED
POTENTIAL GAMES
This study introduces novel gradientbased methods to
enhance exploration in SbPGs which replaces random sam
pling in best response learning Fig 2 highlights the contrast
a Best response learning
b Gradientbased learning
Fig 2 Learning methods during exploration in SbPGs
Fig 3 An example of a 5 5 performance map with a 2D
state space in SbPGs with gradientbased learning
between best response and gradientbased learning Initially
the performance map structure is modified to accommodate the
new approach Instead of storing only the bestexplored action
and its corresponding utility value in each state combination
we now stack selected actions and their utilities for each data
point across various state combinations as shown in Fig 3
Here xandydenote the position in the 2D grid but not
limited to 2D irepresents the index of the relevant player and
pindicates the actual size of the selected actions Additionally
this leads to different updated rules for the performance maps
In the gradientbased learning approach actions are con
sidered as weights to be optimized with respect to the utility
function which serves as the objective function The adjust
ment of actions is guided by the gradient of the utility function
J To promote stable exploration OrnsteinUhlenbeck OU
noise 36 is optionally incorporated which introduces tempo
rally correlated noise that prevents rapid and erratic changes in
actions OU noise also helps in overcoming local optima and
exploring a wider range of states which leads to more robust
learning particularly in continuous action spaces In this study
OU noise can be optionally activated during exploration but
is deactivated during exploitation The impact of OU noise on
our proposed method is further analyzed in our experiments
Therefore actions during exploration are computed as follows
axy
ip1axy
ipα Jγou 8
where axy
iprepresents the previous action in the x y state
combination axy
ip1denotes the new action γousignifies the
OU noise output and αdenotes the learning rate Initially
each action in each data point is set to zero as follows
axy
i0uxy
i0 0 9
The gradientbased learning automatically cancels the update
rules in Eq 3 and 4 By cancelling those gradientbasedlearning achieves faster convergence and smoother exploration
dynamics than bestresponse learning However the learning
rateαmust be properly defined
The challenge is in dealing with nonconvex or unknown
mathematical formulas of utility functions for each player
within SbPGs Players often face unknown mathematical for
mulas of utility functions as these functions are integral to the
system and can be challenging to derive As in Fig 1 the
players interact with the system without knowing their utility
function Instead they learn and optimize their policies based
on the utility values not functions associated with selected
actions in particular states Hence players must internally
estimate the utility function to provide the gradientbased
learner with directional learning signals In response to this
challenge we propose three different estimation variants for
different objective functions systems and complexities These
variants utilize Newtons first divided difference method 37
In the following subsections we discuss the three proposed
variants including the basic estimation method augmented
with momentum and incorporating polynomial interpolation
These three variants operate characteristic estimation function
structures each offering unique advantages The basic esti
mation method calculates gradients directly from the current
iterations utility Momentum enhances this process by inte
grating a portion of the previous iterations gradient to smooth
out fluctuations and accelerate convergence Incorporating
polynomial interpolation takes a step further by utilizing a
polynomial curve to represent historical gradient data which
potentially offers a more subtle and flexible approach to gra
dient estimation Additionally we introduce a kickoff method
aimed at accelerating the training process
A Gradient Ascent with Newtons First Divided Difference