Large language models (LLMs) offer promise for dynamic game content generation, but they face critical barriers, including narrative incoherence and high operational costs. Due to their large size, they are often accessed in the cloud, limiting their application in offline games. Many of these practical issues are solved by pivoting to small language models (SLMs), but existing studies using SLMs have resulted in poor output quality. We propose a strategy of achieving high-quality SLM generation through aggressive fine-tuning on deliberately scoped tasks with narrow context, constrained structure, or both. In short, more difficult tasks require narrower scope and higher specialization to the training corpus. Training data is synthetically generated via a DAG-based approach, grounding models in the specific game world. Such models can form the basis for agentic networks designed around the narratological framework at hand, representing a more practical and robust solution than cloud-dependent LLMs. To validate this approach, we present a proof-of-concept focusing on a single specialized SLM as the fundamental building block. We introduce a minimal RPG loop revolving around rhetorical battles of reputations, powered by this model. We demonstrate that a simple retry-until-success strategy reaches adequate quality (as defined by a rubric-based LLM-as-a-judge scheme) with predictable latency suitable for real-time generation. Generation time estimates based on human annotation and cross-model validation suggest that the retry strategy remains practical even under substantially stricter quality requirements for the quantized models. While local quality assessment remains an open question, our results demonstrate feasibility for real-time generation under typical game engine constraints.

It is an irrefutable fact that the Differential Evolution (DE) algorithm is one of the most widely used stochastic algorithms for solving complex optimization problems. In recent decades, the DE algorithm has garnered significant interest from researchers due to its considerable potential. However, it is evident that the scientific community considers it necessary to modify and create variants of the original version to improve its performance. Within the DE algorithm, the mutation operator has been an essential component in improving performance and exploration/exploitation balance. The present manuscript proposes a novel approach to the mutation operator of the DE algorithm, grounded in the Mahalanobis distance. The Mahalanobis Distance Enhanced Differential Evolution (IMPDE) algorithm is an innovative methodology that aims to explore a new paradigm by measuring the distance of particles to the mean during each iteration. The enhancement of the algorithm is appraised by employing a series of CEC-2017 assessments and traditional optimization problems. Furthermore, a nonparametric Friedman test is conducted to refute the obtained results. This work represents a significant contribution to the ongoing efforts to enhance the optimization process of one of the most prevalent algorithms in recent times.

Maintaining population diversity in evolutionary computation poses a significant challenge throughout the optimization process. A balanced exploration-exploitation dynamic is crucial for achieving optimal results, and a diverse population proportion plays a key role in achieving this balance. Over the years, various methods have been proposed to evaluate diversity. However, relying on a single value to gauge population diversity has its limitations. This paper explores the shortcomings of existing population diversity metrics through extensive experimentation with different population configurations. The findings reveal that some metrics may be misleading, as a lower diversity value does not necessarily imply poor performance in solving optimization problems. In response to these challenges, this paper introduces a new and improved diversity metric, building upon the Dimension-Wise Diversity notion. The proposed metric corrects the shortcomings associated with its predecessors, leading to a more accurate interpretation of diversity values.

This paper proposes a structured methodology to evaluate AI-generated game narratives, leveraging the Delphi study structure with a panel of narrative design experts. Our approach synthesizes story quality dimensions from literature and expert insights, mapping them into the Kano model framework to understand their impact on player satisfaction. The results can inform game developers on prioritizing quality aspects when co-creating game narratives with generative AI.

Differential Evolution (DE) represents an advanced evolutionary algorithm due to its continuous innovation and advances around its operators and applications. In general, the mutation operator is designed to enhance the performance of Differential Evolution (DE) algorithms. The mutation operator is a vital component in enhancing the effectiveness of DE algorithms. The proposed mutation operator, the Lissajous Mutation (LM) is specifically designed to improve the balance between exploration and exploitation through the geometric origin of the Lissajous curves. This paper introduces an up-to-date version of Differential Evolution with Lissajous Mutation, Multi-crossover, and Dynamic selection (IDEL). This new DE variation presents changes in all the fundamental stages, well-known as Mutation, Crossover, and Selection. Through a series of meticulously designed experiments and rigorous comparisons with existing DE variants, the effectiveness of IDEL is demon-strated. IDEL showcases notable improvements in convergence speed, exploration-exploitation equilibrium, and solution quality across diverse benchmark scenarios. This paper contributes to the ongoing growth of DE and presents a novel way to enhance its overall performance.

Particle Swarm Optimization (PSO) has efficiently solved several real-world applications and optimization problems. However, it has shortcomings, such as premature convergence and stagnation at local minima. Inertia weight is a parameter of this algorithm that controls the global and local exploration and exploitation capability by determining the influence of the previous velocity on its current motion. Therefore, this article proposes a PSO with a Diversity-aware Inertia and Velocity Control (PSOIVC) algorithm to improve the PSO performance. The PSOIVC employs a novel diversity-aware inertia weight and velocity control approach to tune the parameters to produce a trade-off between exploration and exploitation of the algorithm using the dimension-wise diversity. The PSOIVC algorithm is compared with eight algorithms, including variants of the PSO, on a set of 30 benchmark functions for a single objective real parameter in 30 and 50 dimensions. Based on the results, the proposal presents significant outcomes according to the average values obtained for both comparisons; because it performed similarly or better than the other algorithms in 23/30 and 16/30 for 30 and 50 dimensions, respectively.

Particle swarm optimization (PSO) is essential to evolutionary computation algorithms (ECA). The PSO has some drawbacks as premature convergence and stagnation at local minima. Inertia weight is a parameter that controls the global and local exploration and exploitation capability in the PSO by determining the influence of the previous velocity on its current motion. This article proposes using a stagnation counter that verifies the times the PSO is stuck in the same fitness value. In the proposed fuzzy controlled PSO with stagnation coefficient (FCPSO), a fuzzy controller is designed to tune the inertia weight based on the population's diversity and the search's stagnation. This modification allows the PSO to escape from suboptimal values enhancing its search capabilities. The FCPSO is tested over 28 benchmark functions in 50 dimensions. Besides, it has been compared with nine optimization algorithms from the state-of-the-art. The experiments and comparisons suggest that the FCPSO is an interesting tool for solving complex optimization problems.

Hyper-heuristics (HH) are strategies that have a high-level mechanism to combine, select or generate heuristics at a low level to find solutions based on the information received during the search process. Typically, an HH approach involves evaluating several algorithms and constructing a priori a structure containing the information required to select, combine or generate the appropriate heuristic to solve a given problem using a usually probabilistic selection method. This paper considers a constructive online learning HH with an agent population selection and control strategy for each heuristic with a reward-punishment approach based on a mean absolute deviation criterion and exploitation-exploration stages to optimize these phases in the optimization process. We compare our proposal, denominated Hyper-heuristic based on the mean absolute deviation metric and the exploration-exploitation stages for the population management (HH-MAD) using uni-modal, multi-modal, composite, and shifted benchmark functions among different optimization algorithms. The experimental results validated the central tendency measures and the non-parametric tests, showing that HH-MAD is competitive, outperforming the other approaches.

Population size is an important variable in evolutionary algorithms (EA). Its proper configuration improves the performance of the search process not only in terms of the fitness function but also for the resources required. This article introduced a population management mechanism that includes different operators. Such operators are designed and applied based on the diversity of the population. In general terms, the operators address problems in EA regarding stagnation and the inefficient use of the function evaluations. As a case of study, the proposed method is applied in the Differential Evolution (DE) to provide it the ability to change its population size according to its needs. The experimental results and comparisons demonstrate greatly improved performance when compared to the unmodified DE, some of its most successful variants, and other much more complex algorithms from the state-of-the-art.

The balance between exploration and exploitation is an important feature in Evolutionary Algorithms (EA). The use of different operators permits to explore the search space and exploit the most prominent regions. This article introduces a dynamic operator selection method that considers different criteria at the same time. The proposed approach uses a dynamic decision matrix (DyDM) to identify which operators must be used at each iteration based on how the algorithm behaves. The DyDM considers specific information as the diversity of the algorithm to avoid stagnation, the actual iteration to work accordingly, and the fitness to direct the search. The proposed approach is called Dynamic Decision Matrix Optimizer (DyDMO) and it has been compared with different well-known algorithms tested on the CEC 2017 benchmark functions. The comparative analysis and non-parametric statistical tests validate how DyDMO im-proves the quality of the solutions and is more stable than its comnetitors.

Successful and accurate modelling of level difficulty is a fundamental component of the operationalisation of player experience as difficulty is one of the most important and commonly used signals for content design and adaptation. In games that feature intermediate milestones, such as completable areas or levels, difficulty is often defined by the probability of completion or completion rate; however, this operationalisation is limited in that it does not describe the behaviour of the player within the area. In this research work, we formalise a model of level difficulty for puzzle games that goes beyond the classical probability of success. We accomplish this by describing the distribution of actions performed within a game level using a parametric statistical model thus creating a richer descriptor of difficulty. The model is fitted and evaluated on a dataset collected from the game Lily's Garden by Tactile Games, and the results of the evaluation show that the it is able to describe and explain difficulty in a vast majority of the levels.

We analyze a series of empirical properties exhibited by customer lifetime value data, including zero-inflation, heavy tails, and the varying behaviour of inter-purchase times and purchase size. Rather than focusing on specific models already established in the literature (e.g., RFM or Pareto/NBD), what is emphasized here are empirical properties of mobile games data which may lead to revisit certain assumptions of existing models.

In this work we investigate whether it is plausible to use the performance of a reinforcement learning (RL) agent to estimate the difficulty measured as the player completion rate of different levels in the mobile puzzle game Lily's Garden.For this purpose we train an RL agent and measure the number of moves required to complete a level. This is then compared to the level completion rate of a large sample of real players.We find that the strongest predictor of player completion rate for a level is the number of moves taken to complete a level of the ~5% best runs of the agent on a given level. A very interesting observation is that, while in absolute terms, the agent is unable to reach human-level performance across all levels, the differences in terms of behaviour between levels are highly correlated to the differences in human behaviour. Thus, despite performing sub-par, it is still possible to use the performance of the agent to estimate, and perhaps further model, player metrics.