Dynamic Difficulty Algorithms as a Tool for Enhancing Player Retention: An Empirical Study in a Gaming Environment

The American Journal of Engineering and Technology

115

https://www.theamericanjournals.com/index.php/tajet

TYPE

Original Research

PAGE NO.

115-123

DOI

10.37547/tajet/Volume07Issue06-12

OPEN ACCESS

SUBMITED

19 April 2025

ACCEPTED

22 May 2025

PUBLISHED

14 June 2025

VOLUME

Vol.07 Issue 06 2025

CITATION

Yurii Sulyma. (2025). Dynamic Difficulty Algorithms as a Tool for Enhancing
Player Retention: An Empirical Study in a Gaming Environment. The
American Journal of Engineering and Technology, 7(06), 115

–

123.

https://doi.org/10.37547/tajet/Volume07Issue06-12.

COPYRIGHT

© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.

Dynamic Difficulty
Algorithms as a Tool for
Enhancing Player
Retention: An Empirical
Study in a Gaming
Environment

Yurii Sulyma

Lead Unity Developer. Cubic Games Kyiv, Ukraine

Abstract:

This article examines the application of

dynamic difficulty algorithms to optimize player
retention and monetization metrics in free-to-play
projects through an empirical study conducted within a
gaming environment. The fact that key indicators of a
projec

t’s viability in the F2P industry, such as D1/D7/D30

retention, directly correlate with LTV and operating
profit, makes the research relevant. Traditional static

difficulty curves give rise to the “difficulty paradox” —

boredom or frustration that accelerates churn. In
contrast, DDA promises to keep the player in

Csíkszentmihályi’s “flow” zone by balancing challenge

and skill. This study aims to demonstrate, on causal data,
the effect of algorithmically adaptive difficulty on user
retention and revenue. The novelty of the work lies in a
large-scale randomized controlled experiment that

combines the segmentation of “at

-

risk” and “core

-

spender” cohorts, as well as an A/B

-testing and RCT

methodology, to evaluate DDA as a scalable product
parameter rather than merely a UX enhancement. The
main findings show that night-by-night decreasing

difficulty for the “at

-

risk” subgroup increases D30

retention by 3 percentage points, yields, on average,
one additional day of play and ten more rounds per user
per month, and an LTV uplift of $ 0.08 per user, where
IAP and 21% by advertising generate 79% of the

increase. The effect is heterogeneous: the “core

-

spender” segment primarily exhibits a financial
response, whereas “frustrated” players increase their

play activity without significant growth in spending. A

The American Journal of Engineering and Technology

116

https://www.theamericanjournals.com/index.php/tajet

comparative analysis revealed that simple heuristics
offer a baseline uplift, while classical ML models can
ensure up to a 20% retention growth. Additionally, RL
agents and hybrid fuzzy-RL solutions can retain players
longer at comparable computational costs. At the same
time, generative LLM-based controllers open up
prospects for unifying DDA approaches. This article will
be helpful to game-product analysts, personalization-
system developers, and monetization managers in the
video-game industry.

Keywords:

dynamic difficulty adjustment; player

retention;

free-to-play;

flow;

algorithmic

personalization; LTV; A/B testing; machine learning.

Introduction:

In the free-to-play model, the viability of a

project is first and foremost measured by retention. The
industry traditionally focuses on three intervals: D1
measures the share of installers who return at least once
the next day, D7 after a week, and D30 after a month.

Historical “pass rates” were approximately 40% / 20% /

10%, but rising user-acquisition costs have shifted the

benchmark: today’s sustainable mobile hits aim for D1 ≈

50% while maintaining former D7 and D30 targets [1].
These first thirty days almost entirely determine
cumulative LTV, since up to 80

–

90% of revenue (IAP +

ads) in casual and mid-core projects is collected within
this interval; AppsFlyer reports an average D90 ARPU of
$3.15 on iOS and $2.15 on Android for Tier-1 markets,
with 65

–

85% of that amount contributed by D30 [2].

Thus, each additional percentage point of early
retention scales revenue nonlinearly: consulting
estimates show that even a 5 pp retention uplift can
boost operating profit by up to 95% thanks to a longer
monetization tail [1].

The main barrier to achieving this uplift is the “difficulty
paradox.” If game challenges remain below the player’s

skill level, boredom ensues; if they sharply exceed it,
frustration occurs, and both states accelerate churn.
Csíkszentmihályi

’s flow model formalizes the skill–

challenge balance: a shift toward low challenge induces
boredom, a shift toward high challenge induces anxiety.
Psychophysiological studies further indicate that low
autonomy

–

induced boredom directly correlates with

increased frustration, amplifying negative affect and
predicting early churn [3]. Developers who rely solely on
static difficulty curves effectively gamble on whether a

predetermined trajectory will fall within the acceptable

“flow corridor” for every new cohor

t.

Dynamic Difficulty Adjustment offers an algorithmic
solution to the paradox. DDA is defined as a system that,
in real time, alters gameplay parameters, scenarios, or
AI behavior based on player telemetry to keep the user
within an optimal challenge zone [4]. Modern
implementations

—

from gradient-based rules to PPO

agents

—

integrate atop the analytics stack and close the

loop “data → churn

-

risk prediction → adaptation,”

thereby transforming retention from a post-hoc KPI into
a controllable product parameter. Therefore, DDA is
now regarded not merely as a UX tool but as a direct
lever for LTV growth, the system extends the active
lifecycle of players by minimizing boredom and
frustration segments. It increases the share of those
valuable D30 users who generate the core profits.

MATERIALS AND METHODOLOGY

The investigation of dynamic difficulty algorithms as a
tool for enhancing player retention is based on the
analysis of 18 sources, including industry reports on
retention and LTV metrics in F2P games [1, 2], review
papers on the concept of Dynamic Difficulty Adjustment
[4], psychophysiological studies of flow states and player
motivation [5, 7], as well as empirical case studies of
machine-learning and deep-learning methods applied in
gaming systems [9, 12]. Additionally, results from
gamification-intervention

meta-analyses

[8]

and

reviews of hybrid fuzzy logic-based approaches [13] and
generative AI controllers [14, 15] were taken into
account.

The methodological framework integrates three key
components. First, a systematic literature review was
conducted, classifying DDA approaches into simple
heuristic rules, classical machine learning models (trees,
boosting,

regressions),

reinforcement

learning

algorithms, and hybrid “fuzzy + RL” solutions [4, 9, 13].

Second, for quantitative evaluation of baseline and
advanced game metrics, data from industry reports by
Solsten and Devtodev on D1/D7/D30 and ARPU/ARPPU
were utilized [1, 2], along with recommendations for
computing LTV via integration of retention curves and
ARPDAU [18]. Third, the methodology for empirical
testing through randomized controlled trials (RCTs) and
A/B tests is described: assigned to either a control

The American Journal of Engineering and Technology

117

https://www.theamericanjournals.com/index.php/tajet

branch with static difficulty or a treatment branch with
adaptive difficulty over 50 days, enabling assessment of
the DDA effect on win probability, progression depth,
D30 retention, and financial metrics [11, 16].

RESULTS AND DISCUSSION

The key psychological mechanism that dynamic

difficulty algorithms seek to sustain is the “flow” state—

an optimal combination of engagement and control that
arises when the subjective challenge of a task matches

the player’s current skills. Csíkszentmihályi’s flow theory

indicates that even a slight mismatch between challenge
and skill shifts the experience into zones of boredom or

anxious frustration; maintaining the balance prolongs
attention, increases enjoyment, and makes a return to
activity more likely [5]. Empirically, this is demonstrated
in a mobile sample. In an experiment with 60 Candy
Crush players, the highest flow scores and strongest
desire to continue playing were observed precisely
when participants tackled just-right levels. In contrast,
overly easy episodes sharply reduced interest and tough
ones provoked increased frustration with only a
moderate rise in flow, as shown in Fig. 1 [6].
Consequently, a DDA system that dynamically aligns the
difficulty curve with individual skill directly reproduces
the condition necessary for the emergence and
maintenance of flow.

Fig. 1. Average boredom, frustration, and arousal scores across the three levels of difficulty [6]

Deci and Ryan’s self

-determination theory provides an

additional explanation for the motivational effect of
DDA. It describes three basic psychological needs

—

competence, autonomy, and relatedness

—

whose

satisfaction enhances intrinsic motivation and supports
long-term persistence. Multicohort video-game studies
show that when interface and content support a sense
of efficacy and freedom of choice, subjective enjoyment
and willingness to return after a session increase; across

four series of experiments, satisfaction of competence
and autonomy needs reliably predicted preference for
continued play and gains in short-term well-being
among players [7]. Adaptive difficulty adjustment
logically fits within this framework because each
successfully overcome challenge confirms competence,
and the ability to influence the difficulty trajectory

through one’s actions sustains a sense of autonomy.

The American Journal of Engineering and Technology

118

https://www.theamericanjournals.com/index.php/tajet

A meta-analysis of 35 gamification interventions
quantitatively shows that mechanics enhancing
autonomy and relatedness produce substantial effects
(Hedges g = 0.638 and 1.776, respectively), whereas the
impact on competence is more modest (g = 0.277) [8].
This indicates that even lightweight gamification
elements can feed key needs, but without proper
calibration of the challenge level, the sense of
competence

remains

limited.

Therefore,

DDA

algorithms are regarded as a missing infrastructural
layer: they simultaneously maintain the skill

–

challenge

balance (flow) and create a sequence of victories and

“skill growth” that addresses precisely the competence

need underserved by classic gamification elements.
Combining both theoretical perspectives, DDA
transforms

basic

psychological

constructs

into

controllable product parameters, explaining its ability to
boost retention over early and mid-game horizons
consistently.

The algorithmic landscape of dynamic difficulty today

lies on a continuum from simple “if–then” rules to

generative models capable of autonomously crafting
challenges for specific users. Such diversity is explained
by the fact that each successive technological wave aims
to uphold the challenge more precisely, the skill balance
described above, thereby more reliably keeping the
player within the flow corridor and satisfying
competence.

At the most basic level, remain heuristic rules: the game
simply tracks several metrics and incrementally shifts
parameters. A classic example is a Difficulty Adjustment
system that changes enemy rank by counting player
damage taken, shot accuracy, and number of retries to
prevent the session from drifting into boredom or
frustration. Methods in this class require minimal data

and computational resources but scale poorly: in live
services with hundreds of levels, designers must
manually author thousands of conditions, and players
with atypical trajectories fall outside the rule set.

The next step comprises classical machine-learning
algorithms. Decision trees, gradient boosting, and
logistic regression predict the probability of failing a flag-
level or the risk of churn and feed these estimates into a

simple difficulty “knob.” In a study based on the Lily’s

Garden puzzle, a neural network trained on a mix of
telemetry and simulated playthroughs achieved the

most stable accuracy in identifying “hard” levels among

ten models, allowing the team to weed out choke-points
before content release [9]. Field A/B tests show that
replacing manual calibration with such ML estimates can
add up to 20% to retention without a noticeable
increase in player frustration [10]. A large RCT involving
300,000 users demonstrated that gently lowering

difficulty for the “at

-

risk” segment increases

engagement and long-term monetization, even
offsetting a short-term drop in IAP [11].

Once sufficient computational power and telemetry
became available, deep-learning methods

—

particularly

reinforcement learning

—

entered the scene. In a MOBA

prototype for League of Legends, DQN, PPO, and TD3
agents analyzed KDA, economy, and map control and
adjusted bot strength every 10

–

20 seconds; the most

advanced configuration increased player satisfaction by
20% and kept them in matches longer at comparable
adjustment intervals [12]. Detailed results of this study
are presented in Fig. 2. Such systems excel because they
learn directly from live interactions. Still, they require
careful action constraints: ill-

considered “tweaks” in PvP

can easily be perceived as unfair.

The American Journal of Engineering and Technology

119

https://www.theamericanjournals.com/index.php/tajet

Fig. 2. Comparison of Key Performance Metrics for RL Difficulty-Adjustment Algorithms [12]

Fuzzy

–

logic

–

based approaches occupy a distinct niche,

and “fuzzy + RL” hybrids. Fuzzy rules describe player
state with terms like “low health,” “high accuracy,” etc.,

then machine-readable if-then statements map into a
space of smooth values. In a shooter prototype, such a
system maintained challenge balance for most
participants without retraining and produced even
progression curves, confirming its suitability for projects
with limited telemetry [13]. These hybrids often serve as
a transitional step for studios that have outgrown pure
heuristics but are not yet ready to invest in heavy RL
training.

The newest layer comprises LLM-based controllers and
other generative AI. Since 2024, prototypes have
emerged in which GPT-like networks generate real-time
tips, restructure quest lines, or even derive new
opponent behaviors. A recent experiment [14] showed

that an LLM trained on simulations of “paradoxical”

games generates strategies that outperform static
templates in profitability and flexibility, confirming the
potential of generative models as a universal DDA layer.
In practice, major vendors are already releasing cloud
SDKs in which the same LLM advises developers on when
and how to nerf a boss. In China, platforms such as

Tencent’s AI Lab attract over 2 million creators monthly,

while India-focused mobile developers use AI to
prototype hyper-casual games. From a gender
perspective, male users (65%) currently outnumber
female creators (30%), though platforms like Pocket

Gems’ Twine and Episode Interactive report an 18%

increase in women since 2022 [15]. Meanwhile, the
global market for AI-based game generators is forecast
to grow from USD 1.8 billion in 2024 to USD 3.72 billion
by 2028, as shown in Fig. 3.

The American Journal of Engineering and Technology

120

https://www.theamericanjournals.com/index.php/tajet

Fig. 3. Global AI Game Generators Market Size Forecast

[15]

In sum, each successive generation of dynamic-difficulty
algorithms enhances personalization: from rigid rules,
through predictive models, to systems that learn and

“think” alongside the player. For the product, this means

increasingly predictable control over key D1/D7/D30
metrics and, consequently, over LTV.

The real impact of adaptive difficulty is almost always
measured via controlled A/B experiments, because only

randomization can isolate the algorithm’s effect from

seasonality, marketing, and heterogeneous cohorts.

Ascarza, Netzer, and Runge’s field study in a popular F2P

puzzle set the industry benchmark: 300,000 players
were randomly assigned to control and adaptive
branches over 12 weeks, yielding statistical power above
0.9 for a 1 pp gain in daily engagement and a 3 pp
increase in D30 retention [16]. After randomization, the
key is selecting appropriate metrics. For event-based
measures

—where “death” is a seven

-day pause

—

Kaplan

–

Meier curves and Cox proportional hazards

models are convenient: they provide an intuitive hazard
ratio and naturally handle right censoring, inevitable in
short tests. If researchers are interested in micro-
behavioral changes before actual churn, mean time to
inactivity (MTTI)

—the average time between a player’s

last activity and the “silence” threshold defined by ad

platforms like Adjust

—

is added to the analytics stack

[17]. The financial effect is measured via ARPU or

ARPDAU, and the integral outcome is captured by the
Lifetime Value metric, for which analytic packages
recommend integrating the retention curve or
approximating via cumulative daily retention [18].

Even with significant aggregate DDA effects, responses
seldom are uniform across all audiences, so the final
analysis stage is segmentation. Such selectivity is
essential for ethical reasons as well. Explaining the

algorithm’s workings down to the bits isn’t necessary.

Still, an explicit “adaptive mode enabled” indicator and

the option to disable it materially mitigate manipulation
concerns, especially in competitive games where
fairness is critical. At the same time, design must avoid

the “trophy paradox”: if victories

come too easily,

competence is devalued and intrinsic motivation drops.
The conflict between transparency and monetization
emerges when reduced difficulty removes the incentive
for pay-to-progress IAPs. Thus, two scenarios arise. In
hyper-casual projects reliant on advertising economics,

DDA acts as a pure “retention

-

lift” tool and almost

always pays off. In mid-core titles built around

“paywalls,” adaptation is applied selectively: it eases

frustration up to the first purchase. Still, it does not
eliminate the need for consumables at gate levels,
where the motivation to pay remains.

The empirical study [11] was designed as a classical RCT:
from June

–

August 2024, 330 000 players who had

The American Journal of Engineering and Technology

121

https://www.theamericanjournals.com/index.php/tajet

passed at least twenty levels and played fewer than 20
rounds in the previous week were randomly assigned
either to a control branch with standard difficulty or to
a DDA algorithm that lowered challenge nightly for

those in the “at

-

risk” group; the share of such treated

users was 41.8%, and assignment remained fixed for all
50 days of observation. This design ensures both
exogenous load distribution and the ability to measure a

long chain of consequences, from the immediate “ease”

of the first session to behavior one month later.

Manipulation checks confirmed that the intervention
indeed eased gameplay: the probability of winning a
round on day 1 increased, and the average score rose by
7,382 points alongside a star gain of +0.297. The
cumulative snapshot at 30 days shows +1 additional
playing day and +10 rounds versus control. Financially,
this translates into an extra $0.08 LTV per user in the first
month; 79% of the uplift comes from IAP and 21% from
advertising due to increased playtime [11].

The authors interpret the observed effect as a

consequence of “accelerated progression”: eased

boards yield a rapid series of victories, elevate the sense
of competence, and reduce frustration risk, thereby
delaying churn and extending the monetization window.
Detailed analysis reveals heterogeneity: players who
already demonstrated a fast level-completion pace
amplify their retention and spending response to DDA;

in contrast, the “frustrated” segment responds mainly

with increased play but almost no rise in spending, while

“core spenders” show modest session changes but

double their revenue uplift, especially when far from the
next gate. Such differentiation confirms that the primary
driver is progress motivation, and economic gains
appear when DDA is targeted at those close to losing
interest yet possessing high purchase potential.

However, the generalization of these results requires
caution. The intervention only involved difficulty
reduction in a single puzzle game; the authors did not
test the symmetric scenario of increasing difficulty for
experts, nor examine potential motivation rebound
beyond the 50-day window. After day 1, adaptation
intensity became endogenous, complicating causal
interpretation of later rounds, and PvP titles may face
fairness concerns if tweaks go unnoticed by opponents.
Finally, the link between retention and LTV is shown in
one F2P economy; the scale effect might differ in games

with strict paywalls. Nevertheless, this work provides
the most compelling empirical evidence that judiciously
targeted dynamic difficulty can simultaneously reduce
frustration, extend user lifetime, and boost marginal
revenue, making DDA a full-fledged product-
personalization tool rather than merely a UX
improvement.

Thus, the empirical study on a representative player
sample confirms that a properly tuned dynamic difficulty
algorithm not only lowers entry barriers and reduces
frustration but also delivers a measurable lift in
retention and monetization. At the same time, the
identified response heterogeneity across cohorts
underscores the need for targeted DDA application and
careful segmentation, especially in games with stringent
payment barriers or competitive contexts.

CONCLUSION

In conclusion, dynamic difficulty adjustment (DDA)
emerges as an effective and reproducible mechanism
for managing key behavioral and financial metrics in
free-to-play (F2P) projects. The experimental design,
involving the randomization of 330,000 players,
demonstrated a statistically significant increase in D30
retention, additional playing volume, and uplift in
cumulative LTV, with most revenue attributable to in-
app purchases rather than advertising monetization.
The results support the central hypothesis that DDA
algorithms, which maintain the challenge

–

skill balance,

indirectly enhance the competence needs satisfaction

and extend the user’s lifecycle.

Csíkszentmihályi’s flow framework and Deci and Ryan’s

self-determination theory explain the observed effect
through maintaining optimal cognitive load and intrinsic
motivation. Algorithmic analysis showed that even a
simple nightly reduction of difficulty

for the “at

-

risk”

segment can trigger a self-

reinforcing “progress spiral,”

in which an early series of victories reduces frustration,
consolidates the sense of efficacy, and consequently
increases the likelihood of reengagement. Cohort
heterogeneity indicated that the groups of users
demonstrating both high spending potential and signs of
imminent churn are the ones that get the most
economic return from targeting. For the "core
spenders," moderate easing is sufficient. However, the
"frustrated" audiences need more interventions beyond

The American Journal of Engineering and Technology

122

https://www.theamericanjournals.com/index.php/tajet

difficulty adaptation.

Limitations pertain to the unidirectional nature of the
intervention (only difficulty reduction), a finite temporal
horizon, and focus on one puzzle mechanic. Long-term
dynamics beyond the experimental period and possible
reputational risks in PvP environments were not
explored either, as was symmetric difficulty increase for
expert players. Nevertheless, this RCT establishes a
robust causal precedent: that a DDA controller can drive
retention growth and margin profit without significant
side effects, given proper segmentation and user
transparency.

Future research directions include broadening the genre
scope to competitive and mid-core titles with paywalls,
comparing algorithm classes

—

from fuzzy-logic/RL

hybrids to generative LLM controllers

—

and analyzing

long-term adaptive difficulty impacts on monetization
model sustainability. Scaling such experiments will help
delineate the applicability boundaries of DDA and
develop industry standards for ethical player
notification,

while

preserving

the

competitive

advantage

of

personalized

challenge-curve

management.

REFERENCES

C. Ansbacher, “The True Drivers Of D1, D7, And D30
Retention In Gaming,”

Solsten

, Dec. 07, 2023.

https://solsten.io/blog/d1-d7-d30-retention-in-gaming

(accessed Apr. 15, 2025).

D. Byshonkov, “Game Market Overview,”

Devtodev

, Jan.

06,

2025.

https://www.devtodev.com/resources/articles/game-
market-overview-the-most-important-reports-
published-in-december-2024

(accessed Apr. 16, 2025).

E. A. J. van Hooft and M. L. M. van Hooff, “The state of
boredom: Frustrating or depressing?”

Motivation and

Emotion

, vol. 42, no. 6, pp. 931

–

946, 2018, doi:

https://doi.org/10.1007/s11031-018-9710-6

.

M. Zohaib, “Dynamic Difficulty Adjustment (DDA) in
Computer Games: A Review,”

Advances in Human-

Computer Interaction

, vol. 2018, pp. 1

–

12, Nov. 2018,

doi:

https://doi.org/10.1155/2018/5681652

.

J. Sutton, “Mihály Csíkszentmihályi: The Father of Flow,”

Positive

Psychology

,

Mar.

31,

2025.

https://positivepsychology.com/mihaly-

csikszentmihalyi-father-of-flow/

(accessed Apr. 17,

2025).

C. J. Larche and M. J. Dixon, “The relationship between

the skill-challenge balance, game expertise, flow and the

urge to keep playing complex mobile games,”

Journal of

Behavioral Addictions

, vol. 9, no. 3, pp. 606

–

616, Oct.

2020, doi:

https://doi.org/10.1556/2006.2020.00070

.

R. M. Ryan, C. S. Rigby, and A. Przybylski, “The

Motivational Pull of Video Games: A Self-Determination
Theory Approa

ch,”

Motivation and Emotion

, vol. 30, no.

4, 2006, doi:

https://doi.org/10.1007/s11031-006-9051-

8

.

L. Li, Khe Foon Hew, and J. Du, “Ga

mification enhances

student intrinsic motivation, perceptions of autonomy
and relatedness, but minimal impact on competency: a
meta-

analysis and systematic review,”

Educational

Technology Research and Development

, vol. 72, Jan.

2024, doi:

https://doi.org/10.1007/s11423-023-10337-

7

.

J. Theiss Kristensen, “Difficulty Modelling in Mobile

Puzzle Games: An Empirical Study on Different Methods

to Combine Player Analytics and Simulated Data,”

Arxiv

,

Jan. 30, 2024.

https://arxiv.org/html/2401.17436v1

(accessed Apr. 18, 2025).

“Game Balancing Tools for Developers

- Perfect Your

Mobile Game Mechanics,”

Mold Stud

, Mar. 17, 2025

—

https://moldstud.com/articles/p-game-balancing-tools-
for-developers-perfect-your-mobile-game-mechanics

(accessed Apr. 2025).

E. Ascarza, O. Netzer, and J. Runge, “Personalized game

design for improved user retention and monetization in

freemium games,”

International Journal of Research in

Marketing

, Jan. 2025, Accessed: Jan. 26, 2025. [Online].

Available:

https://business.columbia.edu/sites/default/files-
efs/citation_file_upload/Personalized_games.pdf

J. Zhang, “Implementation and effect evaluation of

dynamic difficulty adjustment based on reinforcement
learni

ng in Multiplayer Online Battle Arena games,”

Dean

&

Francis

.

https://www.deanfrancispress.com/index.php/te/articl
e/download/2593/TE006072.pdf/10564

(accessed Apr.

20, 2025).

P. D. Paraschos and D. E. Koulouriotis, “Fuzzy Logic

-

Based Dynamic Difficulty Adjustment for Adaptive Game

The American Journal of Engineering and Technology

123

https://www.theamericanjournals.com/index.php/tajet

Environments,”

Electronics

, vol. 14, no. 1, p. 146, Jan.

2025,

doi:

https://doi.org/10.3390/electronics14010146

.

K. H. Cheong, J. Zhao, and T. Wen,

“Adaptive strategy

automation with large language models in paradoxical

games,”

Physical Review Research

, vol. 7, no. 2, Apr.

2025,

doi:

https://doi.org/10.1103/physrevresearch.7.l022012

.

“AI Game Generators Market,”

PW Consulting

Information & Electronics Research Center

, Mar. 21,

2025.

https://pmarketresearch.com/it/ai-game-

generators-market/

(accessed Apr. 22, 2025).

E. Ascarza, O. Netzer, and J. Runge, “Personalized Game

Design for Improved User Retention and Monetization

in Freemium Mobile Games,”

SSRN Electronic Journal

,

2024, doi:

https://doi.org/10.2139/ssrn.4653319

.

“What

is

the

time

of

inactivity?”

Adjust

.

https://www.adjust.com/glossary/time-of-inactivity/

(accessed Apr. 25, 2025).

“Lifetime metric for game analytics,”

Devtodev

, 2020.

https://www.devtodev.com/resources/articles/main-
metrics-lifetime

(accessed Apr. 26, 2025).