Adaptive AI for competitive gaming: particle-swarm-optimized neural network for skill, engagement, and strategic evolution

Introduction

Artificial intelligence has become an integral part of modern gaming, shaping both gameplay and player experiences (Filipović, 2023). AI agents have accomplished incredible achievements, from non-playable characters in open-world games to complex enemies in the strategy-based environment (Armanto et al., 2025; Younes et al., 2023). The most challenging application of AI is fighting games in which agents have to make fast, real-time decisions and show human-like behaviour, adapting to changing opponent strategies (Guo, Thawonmas & Ren, 2024). Games such as Street Fighter, Tekken, and Mortal Kombat demand strategic depth, quick reflexes, and dynamic planning (Lie & Istiono, 2022). Unlike turn-based or strategy-driven games, fighting games serve as promising candidates for developing adaptive AI agents because they require AI agents to make snap decisions that will take into account their prediction of the opponent’s behaviour (Marino et al., 2021; Sabato & De Pascale, 2022).

Traditionally, AI agents for fighting games are designed with a linear difficulty model where the only difference between agents is difficulty, that is, easy, medium or hard (Paraschos & Koulouriotis, 2023). Meanwhile, this can also quickly be done, but makes for an inaccurate experience when playing against human opponents (Martinez-Arellano, Cant & Woods, 2016; Oh, Cho & Kim, 2017). Using the same set of game mechanics, human players can run an extensive gamut of strategies to beat their opponents (Mendonça, Bernardino & Neto, 2015). For example, a player takes a defensive approach by utilizing the character’s abilities to keep a distance, a controlled pace of the match, and eventually win using their abilities. In contrast, another player continues to use abilities to pressure an opponent at a close distance. One of the key aspects of human play is that it is dynamic. Most fighting game AI agents stop evolving, and as a result, the game becomes static and ultimately less fun for players to play as they are aware of the outcome (Paraschos & Koulouriotis, 2023).

There are many challenges with developing truly adaptive AI for fighting games, as fighting games are games of pure instantaneous decision (with an agent selecting the correct action in the order of milliseconds), and the time between rounds is not desirable (Li et al., 2018). The difference between fighting game AI and most turn-based games is that AI for fighting games must analyze what an opponent is doing, determine what is best to attack or defend against, and execute a response in real-time (Ishii et al., 2018). As reinforcement learning (RL) and nature-inspired optimization techniques permit the AI agent to update strategies iteratively, they are very well suited (Mnih et al., 2015; Díaz & Iglesias, 2019). Most AI agents in fighting games have been using finite state machines (FSM), behaviour trees, or N-gram models. These techniques have a range of flexibility and learning capability, but typically with slow adaptation. While FSMs and behaviour trees yield deterministic behaviour for structured gameplay scenarios, these do not evolve with a human strategy (Armanto et al., 2025).

The most challenging part of AI agents for fighting games is stopping them from overfitting to a specific play style (Kim, Park & Yang, 2020). Most deep reinforcement learning (DRL) methods are trained against fixed opponent strategies; this leads to a scenario where the agent learned for training can only beat certain play styles but be beaten by humans playing different styles (Liang & Li, 2022). This challenge can be overcome by an effective adaptive AI that generalizes well on the variety of opponents it learns about and the different playstyles it can iterate on.

The key is balancing that challenge and fairness in an AI agent trying to fight. This is also the problem with an AI opponent: if they are too strong or exploit game mechanics unrealistically, players become frustrated and disengage (Guo, Thawonmas & Ren, 2024). As a result, the AI agent is not overshadowed by its predictability or ease of defeat. Adaptive AI aims to learn to construct a fair, balanced opponent that can deliver a dynamic challenge and is fun to play with in each match.

Conventional DRL agents are quite powerful but require a lot of training, which means they cannot be used on interactive and real-time games like a fighting game (Li et al., 2018; Hu et al., 2024). The common reinforcement learning (RL) models are generally used in an offline phase, or batch-based learning, which is then converted into action, but the resultant model cannot develop online or dynamical capabilities to adapt and to generalize. Such approaches typically make use of experience replay buffers and iterative backpropagation, which in dynamic environments exhibit latency and lack of stability (Liang & Li, 2022). Such delay-oriented processes hinder flexibility and responsiveness in fast-paced fighting games. In addition to this, DRL methods easily overfit to a specific opponent strategy and are afflicted with catastrophic forgetting when the behavior of the opponent changes (Guo, Thawonmas & Ren, 2024). In contrast, the proposed framework uses Particle Swarm Optimization (PSO) as a gradient-free optimization part of the real-time optimization procedures and adjusts the weights of neural networks on-going during gameplay. In contrast with the classic RL paradigms, PSO does not require long convergence times or a calculation of a gradient, which allows fluid and efficient real-time adaptation of the systems in fighting games in a dynamic environment. Subsequently, PSO offers a more appropriate solution to online, continual learning in interactive, real-time situations.

Swarm intelligence is inspired by PSO, which is a population-based optimization algorithm, and the method of such search employs ideas from realizing a collective behaviour of a swarm (Fang & Wong, 2012). Compared to traditional reinforcement learning techniques that require intensive offline training, PSO facilitates continuous optimization and adaptation and thus arises as a good fit when the objective is to achieve real-time gameplay adaptation (Armanto et al., 2025). Natural optimization processes inspire the proposed adaptive AI. The proposed adaptive AI uses PSO-driven weight adaptation of a neural network for real-time strategic learning to reflect human-like behaviour and increase its strategic depth. This research proposes an adaptive AI fighting game agent that is adaptive to the change of human playstyles using a new hybrid learning approach.

The proposed approach combines PSO with a neural network, producing an adaptable AI agent for fighting games tested in Street Fighter. PSO allows the Adaptive AI agent to learn and adapt its strategy in the matches, whereas static rule-based AI agents follow the predefined script. This ensures more engaging and unpredictable gameplay as the AI is wiser in adapting to scenarios of evolving tactics of handling the opponent’s strategies during gameplay. Using this real-time learning ability, the proposed adaptive AI system provides players with a more immersive and complex playing experience for players, increasing the game’s appeal.

The proposed approach also finds an appropriate balance between exploration and exploitation. The proposed adaptive AI search mechanism can apply PSO-driven strategies to find a balance between searching for new methods and improving what is already reasonable. In fighting games, this balance is important because neither the AI agent relies too much on one tactic of use and can still change between many different opponents. The combination of exploration and exploitation serves as an essential element to ensure that the AI can propose a range of changeable gameplay experiences and thus reduce predictability and increase the player’s engagement. The significant contributions of this research are

• An adaptive AI agent for fighting games using neural network optimization, using PSO to execute real-time strategic learning.

• PSO approach that balances exploration and exploitation, strategic diversity, and decrease of predictability.

• Demonstrating real-time adaptability such that the AI agent can dynamically adapt strategies according to the interactions with opponents having different play styles.

The rest of this article is structured as follows: ‘Related Works’ presents related work in fighting game AI and adaptive agents. ‘Methodology’ outlines the method by which the adaptive agent is created. The result analysis is in ‘Results and Discussion’. Lastly, ‘Conclusion’ of the article presents the study’s conclusion and suggestions for future research areas.

Related works

Real-time decision-making and highly challenging action space make AI research a problematic domain for fighting games. Early AI techniques in fighting games are mostly rule-based and FSM, making the game’s behaviours predictable and with certain limitations. Sato et al. (2015) proposed an adaptive fighting game AI, which was developed by switching between multiple rule-based controllers to make the AI more responsive to various in-game situations, but it lacked adaptability. Firdaus et al. (2024) found that the FSM effectively adds combo systems into retro games. The result of this study shows that the responsiveness of character movements can be significantly improved through the use of FSMs in a way that would improve the gameplay mechanics. These studies are mainly aimed at scaling the difficulty (i.e., making the game harder or easier), however, these are not aimed at human alignment, or human engagement. FSMs are deterministic, they are not flexible enough; therefore, hybrid approaches with adaptive AI techniques will be more appropriate (Zhang et al., 2024).

Methodology

This section presents how PSO-enabled weight adjustment on hybrid neural networks is developed for an adaptive fighting game agent, as shown in Fig. 1. The short-term results are enhanced as the long-term results are better, and our approach does this. The objective is to have the agent adapt and learn its strategies dynamically in the context of the player’s actions so that the agent can optimally achieve the gaming experience. Also, this section explains the methodology with brief descriptions of the neural network structure, PSO weight tuning process, state and action encoding methods, and reward calculations.

The environment of the Street Fighter II (based on the BizHawk Application Programming Interface (API)) presents structured state vectors, which are embedded and passed through a Long Short-Term Memory (LSTM) with attention and then dense layers generate Q-values in their output. The choice of course of action is made by an epsilon-greedy policy with decision latency that is not high (usually less than 100 milliseconds). Transitions are stored in a re-playing buffer and a multi-swarm PSO dynamically updates network weights $\theta$ when new states or opponent strategies are recorded. The average PSO update timing of the system was of the order of 2.3 s and action selection was occurring simultaneously.

Our adaptive AI is based on a neural network that extracts game features and sequence-based dependencies. In section A of the Fig. 1, the principal operational components of the neural network are an input embedding and an LSTM. A layer that makes use of an LSTM layer and a multi-head attention module, as well as thoroughly connected (dense) layers.

The game states combine with actions into state and action vectors, which consist of 28 real values for $x_i\in {\mathbb{R}}^{28}$ and 10 real values for $a_j\in {\mathbb{R}}^{10}$. An embedding layer transforms the inputs into a reduced-dimensional continuous space. The embedding function works through the following definition:

(1) $$E(x_i,a_j)=W_e\cdot \left[\begin{array}{c}{x}_i\\ a_j\end{array}\right]+b_e,$$where:

• $W_e\in {\mathbb{R}}^{d\times (28+10)}$ is the embedding weight matrix,

• $b_e\in {\mathbb{R}}^d$ is the embedding bias vector,

• $E(x_i,a_j)\in {\mathbb{R}}^d$ is the resulting embedded representation,

• $d$ is the dimensionality of the embedding space.

An LSTM layer serves to capture the temporal dependencies that occur during gameplay. The hidden state $h_t$ determines the following calculation at time step $t$ as shown in Eq. (2).

(2) $$h_t=\sigma (W_h\cdot \left[\begin{array}{c}E(x_t,a_t)\\ h_{t-1}\end{array}\right]+b_h),$$where:

• $\sigma$ denotes the tanh activation function,

• $W_h$ and $b_h$ are the LSTM weight matrix and bias vector,

• $h_{t-1}$ is the previous hidden state.

Implementing a multi-head attention mechanism helps the network focus on meaningful time steps. The definition of the attention operation stands as follows.

(3) $$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(\mathrm{Q},\mathrm{K},\mathrm{V})=\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left(\frac{\mathrm{Q}{\mathrm{K}}^{\mathrm{T}}}{\sqrt{{\mathrm{d}}_{\mathrm{k}}}}\right)\mathrm{V},$$where:

• Q, K, and V are the query, key, and value matrices derived from the LSTM outputs,

• $d_k$ is the dimensionality of the key vectors.

After the attention layer, a GlobalAveragePooling1D (GAP) layer aggregates the temporal information:

(4) $${\overline{X}}_d=\frac{1}{T}\sum _{t=1}^T{X}_{t,d},d=1,2,\dots ,D,$$

The predictive operation combines T sequence lengths with D features before normalization. The dense networks receive this vector input to generate the final reward prediction outputs:

(5) $$R(s,a)=f(W_d\cdot h_t+b_d),$$with:

• $W_d$ and $b_d$ being the dense layer’s weight matrix and bias,

• $f$ representing the Rectified Linear Unit (ReLU) activation function.

Section B in Fig. 1 provides the current game state and the previous action as input to the neural network. The information gets to the network, which expects an action for the reward in response. The immediate reward prediction in Eq. (5) combines with other state-action predictions through temporal difference learning to create long-term return estimates for decision-making.

In decision-making tasks involving sequence-based data, the proposed adaptive AI has benefited from deep learning to a great extent. For long-term dependencies, capturing decision accuracy in such environments is essential. Advanced architectures are required for feedforward networks because they do not have memory capabilities. To address this challenge, a LSTM network is used with multi-head attention to process sequential data and efficiently select features.

As the game state and action representation is 38-dimensional, the model takes its first feeding of a 38-dimensional game state and action representation and processes it into an embedding layer. We pass it through a single LSTM layer with 50 hidden units to capture long-range dependencies in game sequences. The architecture of the neural network is shown in Table 2. However, the complexity of the model is increased by adding a multi-head attention layer on top of the model with four heads, allowing the model to concentrate on certain past states during the decision-making process selectively. The network consists of three fully connected layers with successively reduced dimensionality of units (150, 80, and 10 neurons, respectively), all using the ReLU activation function. This structured transformation ensures the final prediction is given optimal feature interaction before the final prediction. The reinforcement of learning processes uses the expectations of the expected reward as an output layer, made of a single neuron and a linear activation function, to predict the reward.

Although multi-head attention has many advantages, as claimed in Attention Is All You Need (Vaswani et al., 2017), a significant drawback is that such a single layer of attention cannot account for deep contextual dependencies. Stacking multiple attention layers may improve hierarchical feature learning, but at a higher computational cost, which does not allow it to be used in real time. Having thus balanced model expressiveness with the efficiency of computation, the architecture provides optimal performance in adaptive AI cases.

Results and discussion

This section contrasts the proposed Adaptive AI agent with the state-of-the-art agents as well as human players. The aim is to identify whether PSO-based optimization makes gameplay more engaging, opponents more adaptable, and performance more efficient overall. In order to be able to interpret the results in a transparent manner, important evaluation metrics are stated in the beginning. Win rate and loss rate refer to the rate of matches won or lost against opponents. Action Diversity Score measures the difference between decision sequences in order to measure the level of behavioural unpredictability. Win consistency score measures the frequency with which results are achieved on 10-game trials, with higher scores meaning greater consistency in success. Opponent Adaptation Score grades the ability of the agent to modify the strategy in regard to the changing opponent behaviour based on performance variability, as an opponent’s style varies. Taken together, these indicators enable a multidimensional perspective of adaptive skill and strategic behaviour.

The correlation heatmap in Fig. 2 provides information about the relationships between different performance metrics. The correlation heatmap results are the mean scores of agents at the agent level by calculating means across the subjects. Élő Rating, Glicko-2 Rating, and win consistency score have a positive, strong correlation, meaning that the AI agents have a high skill rating if the nominal strategies are stable. The correlation of 0.96 between the engagement score and opponent adaptation shows a high correlation between AI models that effectively adapt to their opponent and provide more entertaining gameplay. This is congruent with prior research that fosters the idea that player engagement can grow by enhancing players’ adaptability. On the contrary, a −0.97 correlation score between response time and engagement score means that if an AI agent responds too fast, player engagement decreases. This finding is consistent with research such as Zhang et al. (2024), which showed that response time should be balanced with strategic variation. Despite the fact that Pearson correlations are illustrated in Fig. 2, they were calculated using agent-level mean values (N = 6 agents). Since the sample size is very small and inter-metric correlation exists, the magnitude of the correlation must be seen as a descriptive and not an inferential evidence.

The Élő Rating is an established parameter used to measure the progress of the relative skill level of an agent by working out on game outcomes. Élő is also updated by increasing the score of each agent that achieves some unexpectedly positive result and decreasing it after some adverse result, thereby allowing the trajectories of learning to be monitored over prolonged time scales. More recently, Gao et al. (2024) confirmed that Élő-based measures are becoming widely used to measure increases in AI ability in the competitive setting.

In order to complement Élő, we used the Glicko-2 system, which is an expansion of the Élő system, where they added a rating deviation (RD) that models the uncertainty in the estimate of a skill agent. A low RD is indicative of consistent performance, and a high RD indicates unstable or poorly tested behaviour. In turn, Glicko-2 is a reliable tool for the analysis of improving or learning agents or those agents that adjust to new strategies (Li et al., 2024).

In the present work, Élő ratings and Glicko-2 ratings were updated following every game with respect to common parameters. The proposed Adaptive AI already has an Élő of 1,550.32 and a Glicko-2 of 1,548.56 combined, which presents an adequate degree of talent, stability, and trust in game-plays.

The agents from literature, including Zhang et al. (2024) had 1,532.45 and Glicko-2 1,530.12, Halina & Guzdial (2022) had Élő 1,511.67 and Glicko-2 1,509.23, and Li et al. (2024) had Élő 1,508.23 and Glicko-2 1,505.14, and all of these are outperformed by the proposed adaptive AI. Furthermore, the Human-Agent Collaboration model (Gao et al., 2024) with Élő 1,528.98 and Glicko2 1,526.45 also shows competitiveness with proposed adaptive AI and human strategies. The lower uncertainty of adaptive AI in the Glicko-2 system suggests a more stable and consistent performance, implying excellent and consistent gameplay as shown in Fig. 3. Error bars reflect 95% bootstrap confidence intervals from 1,000 resamples. The human player’s Élő and Glicko-2 ratings are 1,500.00, and the Glicko-2 rating is also 1,500.00. The challenge of outperforming human players compared to those who do, on average, perform a moderate degree of consistency. The Adaptive AI’s Élő rating of 1,550.32 and its Glicko-2 rating of 1,548.56 exceeded these human player averages, not only in its skill level but also in terms of consistency and consistency in performance. This further solidifies the competitiveness of the proposed adaptive AI because, despite its high performance and confidence level, it can still hold its own against the average human player. To statistically prove the fairness of the performance of Adaptive AI over other baseline agents, as well as human contestants, the Wilcoxon signed-rank test was carried out on data summarized in ‘Results and Discussion’. the difference in per-participant Élő deltas of the Adaptive AI and the largest baseline agent was compared using a Wilcoxon signed-rank test. The unit of analysis consisted of individual participants ( $N=60$ pairs). The outcomes showed that there was a significant difference ( $W=15.0$, $p=0.031$). The matched-pairs rank-biserial effect size was $r=0.42$, which is a medium effect.

The output (W = 15.0, $p$ = 0.031) demonstrates that the Élő score of the proposed Adaptive AI is much greater than those of other AI models and the average of humans with a 95 $\mathrm{\%}$ confidence level. This serves to substantiate the idea that the proposed Adaptive AI invariably stands out above the traditional systems with regard to skillfulness and versatility. A measure of the variability in agent skill estimates was obtained by building 95% confidence intervals (CIs) on Élő ratings through bootstrapped sampling. On each of the AI agents, 1,000 random replacement resamples of the outcome in the matches were drawn out of the overall set of matches, with recalculation of Élő scores in each resample. The outcomes of the CIs are illustrated as error bars in Fig. 3. These intervals provide a statistical measure of certainty in the rating of each agent and, in addition, support the fact that the proposed Adaptive AI is consistent in its performance.

In the custom made measures of engagement and adaptation, we provide the 95% confidence intervals, acquired through bootstrap resampling with 1,000 replicates, which we represent in Figs. 4 and 5, in addition to the mean values to prevent inaccuracies of over-confidence through individual comparison. With the engagement score, we measure the deep and varied strategies based on the proposed adaptive AI’s decision-making during gameplay, shown in Fig. 4. Engagement score is a composite, 0–1 scale measure that reflects the ability of an agent to maintain the interest of a user by adjusting its behavior dynamically as the interest of a user wavers. It combines three fundamental aspects, which are variance in match duration, diversity in action, and time randomness. These dimensions are already proven to be strongly connected to the user experience of engagement, as these dimensions regulate perceived responsiveness, novelty, and level of challenge offered by an agent. The formulation can be associated with the visions of human-agent interaction as in Gao et al. (2024) with its focus on understanding outcomes over time and interactivity as critical in engagement, and Halina & Guzdial (2022) with its focus on diversity-driven learning as a process that induces immersion in players, increasing the behavioral variety among states. Normalization to the range of scores between 0 and 1 makes it possible to compare the scores directly across agents, and higher values indicate a more unpredictable, strategic, and human-like play. The proposed adaptive AI is associated with an engagement score of 0.95, which left the interactions during gameplay to closely align with the human verge of 0.98. This result highlights the ability of the system to produce game-play experiences of variation, immersion, and entertainment.

The Shūkai system (Zhang et al., 2024) scored 0.92, and the human-agent collaboration model (Gao et al., 2024) 0.90 also achieved high engagement secondary to dynamic learning, interactive gameplay, and, consequently, higher player involvement. Although FightLadder’s (Li et al., 2024) engagement score was only 0.8, it was lower than the average, primarily due to its less adaptive strategy and set of tactics, which were not as flexible. The Brisket model (Halina & Guzdial, 2022), which has an engagement score of 0.88, demonstrates moderate engagement through diversity-based reinforcement learning but is short of adapting to the chosen adaptive AI. The proposed adaptive AI’s dynamic decision-making and strategic variation contextually replicate a very engaging and unpredictable human gameplay, among the most engaging and enjoyable overall gameplay experiences among all the agents tested.

The opponent adaptation score is the coefficient that shows how fast the AI can adapt its model when new opponent play are introduced, i.e., the adaptability of the AI. The high score of 0.91 means that proposed adaptive AI learns fast from new and old opponents, as shown in Fig. 5. This demonstrates that the proposed adaptive AI has been able to generalize what it has learned from gameplay experience and evolve strategies, which is very applicable in dynamic and competitive environments. The Shūkai system (Zhang et al., 2024) scored 0.89 in adapting to the new challenges and opponents, which means that the adaptation score achieved by the Shūkai system demonstrates its ability to adapt to new challenges and opponents. The human-agent collaboration model (Gao et al., 2024) scored 0.87, showing the effectiveness of hybrid learning approaches in improving adaptability. On the other hand, in comparison to the FightLadder Benchmark (Li et al., 2024), the FightLadder Benchmark received a much lower score of adaptation of 0.77, which indicates it uses a more fixed strategy and less real-time learning and adaptation. The Brisket model (Halina & Guzdial, 2022), with a score of 0.85, is found to be moderately adapted due to diversity-based learning, which contributes to improved adaptability of the AI but is not as efficient as the reinforcement learning-based model in the adaptive AI. Overall, the results show that proposed adaptive AI can withstand and treat opponent adaptability on the side of the opponent and even perform well in competitive and changing environments.

Consistency score shows how well the AI does to win, irrespective of the match scenario, and the win consistency score measures the consistency it can deliver. This metric is crucial if we want to understand the robustness of an AI agent in real-life uses. A similar kind of evaluation has also been carried out in studies such as Li et al. (2024), emphasizing consistency’s importance in robust AI performance. The AI’s ability to learn and adapt to new opponents is also tested using the opponent adaptation score, which measures the AI’s efficiency in changing tactics to adapt and become more generalized. Such an approach is consistent with recent studies like (Halina & Guzdial, 2022) among others, and assessed that adaptability to opponent behaviour is a key factor for AI performance evaluation. The proposed adaptive AI surpasses all other AI models except the human player, with an adaptation score of 0.91. It can modify the strategies dynamically, making it more successful in competitive matches. By demonstrating that the proposed adaptive AI achieves a win consistency score of 0.88, a similar result as (Zhang et al., 2024) 0.88 and (Gao et al., 2024) 0.86, which means that it has consistently high win rates as shown in Table 5.

The response time metric measures the reaction time of AI agents in the decision-making process in environments such as fighting games, where speed matters the most for real-time performance. As stated in previous work, such as Zhang et al. (2024), games depend on the quickness of the AI to respond to in-game events; this metric is highly relevant to games.

Regarding response time, the proposed adaptive AI results in 2.3 s, although higher than other models in the literature. With this quick response time, the AI can make decisions on time in a real-time game experience, which enhances performance in different dynamic environments where the reaction speed is significant to the outcome. However, the response time of the Shūkai system (Zhang et al., 2024) is 2.4 s, which is 0.1 s slower than the proposed adaptive AI. A response time of 2.6 s for the human-agent collaboration model (Gao et al., 2024) reveals that adopting human strategies might incur minor delays in decision-making. With a response time of 2.7 s, the brisket model (Halina & Guzdial, 2022) is slower than the proposed Adaptive AI and Shūkai system, and the FightLadder benchmark (Li et al., 2024) response time is 2.9 s. These results show that proposed adaptive AI still has a slightly faster response time; it makes decisions quickly and performs best in a high-speed and real-time gaming environment. This enables it to be highly effective and competitive at a high level of strategy and execution speed.

To determine the internal validity of the proposed Adaptive AI, we used ablation experiments where key parts of the architecture were eliminated or changed. Table 6 shows the impact of such adjustments on all the evaluation metrics applied in Table 5

Table 6 shows that in all the conditions, no ablation variants had a higher performance compared to the full Adaptive AI. The removal of online PSO weight updates (“No PSO’) resulted in the sharpest decreases in rating systems and custom metrics, but the response time was reduced since no optimisation overhead was present. Maximizing optimization to one swarm also decreased Élő and Glicko-2 scores as well as adaptation and consistency, indicating the stabilizing impact of multi-swarm diversity. When attention-enhanced LSTM was substituted by vanilla LSTM, the engagement and consistency decreased significantly with adaptation decreasing as well (though not as low as the full model), which proves the importance of attention in capturing long-term dependencies. Lastly, the sensitivity analysis was performed by adjusting the PSO coefficients ( $c1,c2$ between 1.2 and 2.0) and the weight of inertia ( $w$ between 0.6 and 1.0) and keeping the multi-swarm and attention set-up. There was performance lowering at more extreme values but was always below the reported set of tuned full model ( $c1=1.5$, $c2=1.7$, $w=0.8$ initial) indicating that the reported settings were a near-optimal.

These findings substantiate the conclusion that the good performance of the adaptive AI is attributed to the synergistic effect of the PSO-based online learning system, multi-swarm optimization, and attention-enhanced sequence modeling, and that the given parameter setup gives the best trade-off between adaptability, engagement, and consistency.

It turns out that the adaptive AI consistently performs at the top for these metrics more often. The proposed adaptive AI performance trend plot results in Fig. 6 show that adaptive AI always maintains high metrics scores, proving that its natural optimization-based approach is sufficient to sustain high performance in competitive duties. For instance, some models, including the FightLadder Benchmark (Li et al., 2024), are notable for decreasing performance due to reliance on their static strategies. On the other hand, brisket (Halina & Guzdial, 2022), human-agent collaboration (Gao et al., 2024), in contrast to those, have the performance increase in the models through their dynamic learning and opponent interaction. It is straightforward across different metrics how consistently adaptive AI performs highly, especially considering it can adapt and evolve its strategies over time, making it remain competitive against AI and human opponents. In the case of engagement, adaptation, and consistency measures, bootstrap resampling (1,000 replicates) was used to estimate the 95% confidence interval around the means to measure the uncertainty of the mean value. To prevent over-precision due to comparison of single values, these intervals are indicated in Figs. 4–6. The results show that the proposed adaptive AI model performs better than the other AI approaches in all key metrics. The proposed adaptive AI uses the dynamic strategy of adapting behaviour to play high-performance with diverse behaviour across different match scenarios. The proposed adaptive AI is a new standard for AI-driven opponents in fighting games, performing very well regarding the agility and speed taken in decisions, with a good balance between making decisions and various attack methods. In general, these findings support that adaptive learning models can bridge the gap between human and AI performance and thus would make gameplay more immersive, interactive, and diversely strategic.

The significance of this work is not related only to the particular case of Street Fighter II. By showing to the world that the online adaptation of the neural network weights can be facilitated using gradient-free, multi-swarm PSO, the research points to the overall mechanism of adaptive decision-making in dynamic settings. The proposed approach merges adaptability and interaction at the same time being computationally efficient in contrast to the earlier adaptive game AIs where the difficulty is limited only to its hardest possible mode or the visions are human-driven. This makes the approach a contender that makes applicable steps to further studies such as training bots to play various games with competitive rule setbacks, adaptive agents to simulation-based education, and scalable multi-agent coordination issues. The results introduced in the article can be generalized to other franchises and areas in further research with the help of the reproducibility package and thus support the originality and the applicability of the method.

Even though it was found that adaptive AI with the assistance of PSO is efficient in the context of a fighting game, it is necessary not to ignore certain drawbacks. First of all, the research is domain specific: all experiments were conducted within the framework of the game of a genre called Street Fighter II thus it is still unknown whether the strategy is generalized to other genres of games or more complex multimodal settings. Second, this state representation was done using engineered features which are exposed by emulator API rather than raw sensory inputs, which has contributed to reproducibility, but not efficiently reflective of the complexity of end-to-end perception pipelines. Third, the human experiment was not justified on the blinded identification, i.e. the subjects were aware of the fact that the experiment was playing against the AI or the baseline model, a possibility that might influence the reaction of the subjects. Future work versions will further scale up to provide whether more or less supplementary game franchises, accommodate additional end-to-end perceptual inputs with convolutional or vision-transformer encoders, and blinded or double-blind experiment environments as solutions to reduce any bias in human-agent evaluation.

Conclusion

We proposed an engaging, strategic, decision-making, adaptive AI for fighting games using PSO-based Neural Networks (NN). With efficient weight optimization by PSO, the proposed model enhances learning stability and convergence, attains faster adaptation to the opponent’s strategies, and keeps performance consistency. The adaptive AI is evaluated through a comprehensive analysis of skill proficiency, after the engagement, and real-time response efficiency, and its performance is superior to that of all existing AI agents and human benchmarks. The proposed model is verified experimentally, achieving an Élő rating of 1,550.32 and a Glicko-2 score of 1,548.56, which are very high skill levels and show performance stability. The engagement score (0.95) is very close to human players (0.98) and superior to human agent collaboration (0.90) and the Shūkai (0.92) alternative models. Additionally, the opponent adaptation score (0.91) also demonstrates that AI is more adaptable than Brisket (0.85) and FightLadder (0.77), which use more rigid predefined strategies. Moreover, the proposed agent can make real-time tactical decisions in a fast-paced, competitive environment with a response time of 2.3 s, which is a key factor. The results show that the NN-based adaptation mechanism combined with PSO enhancement can be a robust replacement for traditional reinforcement learning, yielding high skill retention, fast ability learning, and good response efficiency. These results suggest that such integration of AI opponents driven by PSO will yield engaging gameplay to the player and aid in long-term player retention. The aim of the future research will be to build on the current work by studying the concept of scalability over heterogeneous interactions among agents, including hybrid optimization methodologies and domain adaptation, to deepen the in-game experience of AI players. In the present research, the participants were not blinded specifically to the identities of AI, but the match order was mixed, and the consequences of the perceptual bias were considered. Further studies will involve more drastic control conditions to provide more empirical support for the process of human-AI interaction.

Appendix

Metric definitions (formulas and parameters)

Supplemental Information