Dynamic difficulty adjustment

Dynamic difficulty adjustment (DDA) is the automatic, real-time modification of a game's features, behaviors and scenarios according to the player's skill, so that the game does not become boring because it is too easy or frustrating because it is too hard. In traditional games, difficulty rises linearly or in steps during play, and features such as frequency, starting levels or rates can be set only at the beginning, by choosing a difficulty level. Other names include dynamic difficulty adaptation and dynamic game difficulty balancing, which Fuentes Perez and colleagues define as "a process that adjusts gameplay parameters in real-time according to the current player skill level".

DDA systems estimate the challenge a player faces from game performance, from the player's own ratings or from physiological signals such as heart rate and brain activity, and then change game parameters such as the behavior of opponents or the layout of levels. Methods range from hand-written rules and probabilistic optimization to machine-learning models of player experience and reinforcement learning.

In virtual reality (VR), DDA has been studied for exergames, rehabilitation and training, where the challenge to be adjusted can be physical as well as cognitive. Writing in 2021, Huber and colleagues observed that Beat Saber "only uses predefined sequences and does not dynamically adjust the difficulty", and that the few VR exergames that used DDA at that time tuned only a small number of in-game parameters with heuristics.

Reviewed 1 November 2026. Citation records, quotes, figures, dates and attributions checked against the full texts of Hamlet, the Booth slides and the Zohaib, Huber, Kamikokuryo, Mantz, Pedersen and Li papers, and against the abstracts of the other cited papers About review dates.

How it works

According to Zohaib's 2018 review, all DDA methods share a requirement to measure, explicitly or implicitly, how difficult the game is for the player at a given moment. The review calls these measures heuristic or challenge functions; typical examples are the success rate of hits, the numbers of pieces won and lost, life points, the time taken to complete a task, or other score metrics. Kamikokuryo and colleagues likewise write that most DDA approaches adjust the game according to "the ratio of gains to losses, the time required to complete a specific task, and the remaining life points". The adjustment itself can change the supply of health and ammunition or the accuracy of enemies, the speed, size and distance of targets, the layout of procedurally generated levels or a game timer.

Zohaib's review relates DDA to Csikszentmihalyi's "flow channel": players kept away from boredom and frustration travel through this channel, in which difficulty can rise gradually because players have enough time to learn and improve. Koster applied the concept to games in A Theory of Fun for Game Design. Robin Hunicke and Vernell Chapman likewise described the aim of their Hamlet system as keeping the player in the flow channel, away from states where the game is far too challenging or far too easy. Andrade and colleagues, as summarized in the same review, proposed three requirements: the game must track the player's ability and adapt to it rapidly, follow the player's improving or falling level, and adapt in a way that players do not clearly perceive, with each new game state coherent with the earlier ones.

Reactive and proactive adjustment

Hunicke and Chapman's 2004 paper on Hamlet described two kinds of adjustment action that the completed system would support. Reactive actions adjust elements that are "in play", that is, entities that have noticed the player and are attacking, through the accuracy or damage of attacks, the strength of weapons or the level of health. Proactive actions adjust "off stage" entities that are inactive or waiting to spawn, through their type, spawning order, health, accuracy, damage and packing properties. The authors described the choice as a trade-off. Reactive adjustments are simple to execute and map more directly to changes in the gameplay experience, but they "run the risk of disrupting the player's sense of disbelief" and can make the immediate behavior of in-game obstacles difficult to interpret. Proactive changes happen "in the wings" and are less likely to interrupt the player's suspension of disbelief, but because they act at a greater distance from the point of action they are harder to evaluate, and in the worst case they may require subsequent reactive adjustments, "causing a spiraling loop of change".

A systematic review of personalized VR by Li, Zhu, Liang and Wang, an arXiv preprint first posted in October 2025 and revised in July 2026, applies the same two terms to whole systems: reactive adaptation acts after a deviation is detected, while proactive adaptation predicts future conditions and adjusts before problems arise. The review found reactive adaptation to be the dominant approach across the literature it covered, particularly in gaming and immersive interaction, where adjustments such as difficulty changes are triggered by measurable signals such as performance or behavioral input. It notes that because such adaptations occur only after an event has emerged, they "may disrupt UX in highly immersive or time-sensitive scenarios", and that proactive adaptation remains less explored because systems struggle to predict user states, intentions and environmental changes in real time.

Inputs

Besides in-game performance, some systems ask the player. In Huber and colleagues' VR exergame, players rated each generated maze on a five-point scale from "not at all difficult" to "extremely difficult", and the gap between that rating and the target level of 3 was given to the learning algorithm as a negative reward; an electrocardiogram (ECG) was recorded separately to measure exertion for the evaluation. Other systems use physiological signals, including heart rate in VR exergames, electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS).

Headset hardware affects which signals are practical. According to the Li review, eye tracking built into headsets such as the HTC Vive Pro Eye, Meta Quest Pro and Apple Vision Pro has lowered the barrier to adopting eye measurement in VR systems. EEG devices, by contrast, usually need their own head-mounted setup, which "may conflict with wearing VR HMDs simultaneously", and setups that combine EEG with a VR head-mounted display are typically confined to controlled experiments.

History

Hamlet, described by Hunicke and Chapman of Northwestern University in a 2004 AAAI workshop paper, was a DDA system built on Valve's Half-Life game engine. It used techniques from inventory theory and operations research to analyze and adjust the supply and demand of game inventory, which in a first-person shooter includes health, weapons, ammunition and power-ups. The paper started from the observation that "Video Games are boring when they are too easy and frustrating when they are too hard", and that although most single-player games let players choose a basic difficulty, their overall level of challenge was "often static in the face of individual player input". Hamlet modeled the total damage a player takes as a Gaussian distribution in order to predict when the player was "flailing", which the authors define as "repeatedly edging towards a state where her current means can no longer accomplish necessary and immediate ends". It could then respond by, for example, placing a health pack somewhere in the scene or reducing the accuracy or strength of enemy attacks. Its example policies were a "Comfort Zone", which aimed to keep the player at about 50% health, and a "Discomfort Zone" for more experienced players, who would be "fighting back from 15 or 20% health most of the time". The authors noted that "DDA systems take control away from the designer and put it in the hands of code" and wrote: "Few commercial developers have implemented adjustment systems for their games, and even fewer have shipped them." They planned to evaluate the system by attaching a heart-rate monitor to players who would play the game on both settings. In a 2005 paper, Hunicke wrote that, according to conventional wisdom, players will feel "cheated" if games are adjusted during or across play sessions, and presented preliminary evaluation results for Hamlet that "challenge common assumptions about player enjoyment and adjustment dynamics".

A related system in a commercial game adjusted pacing rather than difficulty. In 2009 presentation slides on the cooperative survival-horror game Left 4 Dead, Michael Booth of Valve described the "Adaptive Dramatic Pacing algorithm" of its "AI Director", which estimated the "emotional intensity" of each survivor, raised for example by injury or incapacitation, and removed major threats for a while when the highest intensity in the team was too high; otherwise it created "an interesting population of threats". Booth summarized: "Algorithm adjusts pacing, not difficulty".

Zohaib's review tabulated the DDA studies it covered by year: 4 in 2009, 13 in 2012 and 14 in 2017, the highest yearly count, which the author described as a significant increase over the decade.

Methods

Zohaib's review groups DDA approaches into probabilistic methods, single and multilayered perceptrons, dynamic scripting, the Hamlet system, reinforcement learning, upper confidence bound for trees (UCT) combined with artificial neural networks, self-organizing systems with artificial neural networks, and affective modeling using EEG. The table summarizes these families together with approaches used in later work.

Approach	How it works	Examples
Rules and heuristics	A challenge function is compared with thresholds or predictions, and hand-written rules change game parameters.	Hamlet predicted inventory shortfalls in a first-person shooter and adjusted items and enemies. In a game that tracks scores, adjustment can be triggered when the score difference between players exceeds a threshold. According to Huber and colleagues, many non-VR rehabilitation exergames adjust a small number of parameters, such as the speed of in-game objects, with heuristics based on the player's performance.
Probabilistic optimization	The player's progression is modeled as a probabilistic graph, and difficulty is set by optimizing its transition probabilities for an objective.	Xue and colleagues maximized players' stay time in level-based games and reported applying the technique in multiple games by Electronic Arts, with "up to 9% improvement in player engagement with a neutral impact on monetization".
Dynamic scripting	Each opponent type has a rulebase. Rules that led to success gain weight and rules that led to failure lose weight, with the sum of all weights kept constant, so that the opponents' tactics match the player's level of experience.	Zohaib describes three variants for balanced play (high-fitness penalizing, weight clipping and top culling) and reports that the first was not successful while the other two did well.
Player-experience models	A neural network is trained on gameplay data and player reports to predict how players will experience content.	Pedersen, Togelius and Yannakakis used evolutionary preference learning to train neural network models on data from 480 sessions of a modified version of the platform game Infinite Mario Bros. A single-neuron model predicted reported challenge and frustration with accuracies of 77.77% and 88.66%; the 69.18% accuracy for fun suggested that more complex non-linear models were needed.
Reinforcement learning and evolutionary computation	An agent learns from rewards which adjustments or opponent behaviors keep play balanced.	Tan, Tan and Tay presented two adaptive algorithms that use ideas from reinforcement learning and evolutionary computation to scale the difficulty of the game AI during play. According to Zohaib's review, both kept score differences at 4 or less in at least 70.22% of games against static computer opponents. Sekhavat proposed multiple-periodic reinforcement learning (MPRL) for a rehabilitation game. Huber and colleagues trained a network with the Deep Q-learning algorithm to build exergame levels.
Multi-armed bandits	The agent picks the next task from observed rewards. Kamikokuryo and colleagues describe the bandit as a specific form of reinforcement learning that explores and exploits an environment without changing its state and does not need large amounts of data.	Kamikokuryo and colleagues compared Boltzmann and sibling Kalman filter exploration for choosing rehabilitation hand movements in VR.
Tree search with neural networks	Adjusting the UCT simulation time produces opponents of increasing difficulty; a neural network trained offline on data generated with different simulation times bypasses the computational cost of the search.	According to Zohaib's review, Li and colleagues used the approach for the ghosts in a Pac-Man test bed; the review notes that UCT itself is not practical for complex online games because it is computationally intensive.
Fuzzy cognitive maps	Game variables are concepts linked by fuzzy and probabilistic causal relationships that can be updated in real time.	Fuentes Perez and colleagues ran simulation experiments with evolutionary fuzzy cognitive maps in a runner-type game.
Procedural content generation	New levels or level segments are generated to match an estimate of the player's ability.	According to Zohaib's review, Polymorph, by Jennings-Teats, Smith and Wardrip-Fruin, used multilayer perceptrons trained on play traces to choose level segments in a 2D platformer. Huber and colleagues generated VR exergame mazes that adapted the required cognitive and physical effort together.
Physiological and affective signals	A model estimates the player's state from body signals, and the game adjusts when the estimate leaves a target range.	According to Zohaib's review, Stein and colleagues triggered adjustment when short-term excitement estimated from EEG dropped below a threshold; in a customized version of the Boot Camp game, players preferred this to a heuristic method and to no adjustment. In a study titled "Dynamic difficulty using brain metrics of workload", Afergan and colleagues used fNIRS to add or remove unmanned aerial vehicles in a path-planning simulation and reported 35% less error than a baseline condition. VR exergame studies have used heart-rate zones.

Use in VR games

Huber and colleagues argue that exergames are harder to adjust than traditional games because physical effort and level complexity both have to fit the player: a level that is too complex but physically appropriate should change only in complexity, and the reverse. Their prototype, tested with an HTC Vive Pro, generated mazes whose walls the player could not see over and whose rooms required exercises designed to prevent back pain; a deep reinforcement learning model built each new maze from the player's rating of the previous one. Headset play limited the choice of exercises. Lying down was excluded because it is "very hard to lay down or get up while wearing an HMD and holding controllers in each hand", and foot-specific movements were excluded because a basic VR setup cannot track them. The game checked each exercise by testing whether the controllers followed a predefined path, and varied exertion through the number of repetitions.

The Li review describes early work on VR games as focused mainly on reactive adaptation, such as difficulty adjustment based on player performance, aimed at maintaining engagement and preventing frustration; more recent work adds predictive and generative techniques for proactive personalization. It also notes that the immersive nature of VR introduces cybersickness, and that the greater complexity of 3D interaction and virtual environments leads to more sophisticated systems and high-dimensional data streams.

Mantz and colleagues studied DDA for competitive two-player VR exergames, in which differences in fitness can lead to anxiety, fear of failure or frustration. After focus groups in which participants watched or played Beat Saber, they built a rhythm game in Unity 2021.3.16f1, based on open-source code that simulates Beat Saber gameplay, with a two-player mode in which the players see each other as avatars. Movement fills each player's bonus bar, and the game adds a factor based on the score difference between the players and the mistakes each has made, so the bar of the player who is behind fills faster. A full bar spawns a power-up block and an attack block at the same time on the left and right sides, so the player has to decide which one to hit. A power-up makes the game easier for the player and an attack makes it harder for the opponent, in each case for 10 seconds, and the strength of a bonus also depends on the score difference. After playtests, the designers added an upper threshold to the fill factor to avoid too many bonus spawns. The adaptation is explicit: players can see that they receive more bonus blocks.

A related approach adapts content to the player's emotional response instead of skill. A method for adaptive VR horror games presented by de Lima, Silva and Galam uses machine learning to identify, from gameplay data alone, which horror elements an individual player fears, and an agent-based system that tracks the horror intensity the player experiences and moderates the use of those elements to intensify fear.

VR rehabilitation and training

Kamikokuryo and colleagues note that patients may become bored if a rehabilitation game is too easy or frustrated if it is too difficult, and that providing an accurate level of challenge during therapy "can be a difficult and time-consuming task for healthcare professionals", because it requires continuous analysis of data from earlier and new rehabilitation sessions. They report that most serious VR games use a static difficulty, selected manually, or a "staircase" progression that raises the difficulty after the player completes set tasks, and that predefined levels let players choose an inappropriate level and do not automatically account for improvement. In their method, hand movements drawn with an Oculus Rift S controller are compressed by an adversarial autoencoder into a two-dimensional latent space. Each of six bandit arms stands for one movement, and a larger distance from that movement's target in the latent space raises its probability, so the arms carry information about how difficult each movement is for the patient. The authors chose a bandit over a full reinforcement learning algorithm such as Q-learning because the states and transition probabilities of a Markov decision process could not be determined at that stage.

The same paper describes Sekhavat's multiple-periodic reinforcement learning (MPRL) approach for a rehabilitation game in which the player's movement, recorded with a Kinect, controls the arm of a virtual character that must hit balls placed on an arc; the learning agent increases or decreases the speed, size and distance of the balls according to whether the player succeeds or fails in hitting the ball. Sekhavat reported that MPRL outperformed traditional multiple-objective reinforcement learning (MORL) in user satisfaction and in improving the movement skills of patients.

For training, Dey, Chatburn and Mark Billinghurst built a cognitively adaptive VR system in which the complexity of a target-selection task followed task-related alpha activity measured with a 32-channel mobile EEG system. According to the Li review, the system increased the task difficulty when the cognitive load indicated by alpha activity was too low, and lowered it when the load was too high. A 2020 systematic review of adaptive VR-based training by Zahabi and Abdul Razak, which searched Compendex, Web of Science and Google Scholar, proposed a framework of performance measures, adaptive logic and adaptive variables. It concluded that such training could be improved by using real-time kinematic or kinetic data and physiological measures, offline measures such as the trainee's profile, adaptations on controlled elements of the simulation, adjusted feedback content, type and timing, and reinforcement learning algorithms, and that these recommendations still needed validation in longitudinal studies comparing adaptive and non-adaptive training.

The Li review found that VR training systems still rely mainly on reactive strategies because these give predictable performance and transparency to instructors and students. It describes proactive adaptation as relatively rare in healthcare VR, where clinical safety and precision come first and the risk of adverse side effects leads systems to choose transparent, reactive frameworks.

Research

The table lists VR studies of difficulty adjustment and related adaptation, with participant numbers, conditions and results as reported by their authors.

Study	Application	Input	What was adapted	Participants	Reported results
Dey, Chatburn and Billinghurst, IEEE VR 2019	Cognitive training: selecting target objects in VR	Task-related alpha activity from a 32-channel mobile EEG system	Task complexity, over 20 levels	14	Significant differences in brain activity as task complexity increased; response time did not change with task difficulty
Huber and colleagues, IEEE SSCI 2021 (arXiv version)	Exergame: maze with exercise rooms, HTC Vive Pro	Player's difficulty rating after each maze	Maze layout and exercise rooms, generated with deep reinforcement learning	19 students	For 13 players who found the first maze too easy or too hard, the unified in-game rating moved from 1.69 to 2.46 toward the target of 3; mean heart rate rose from 89.96 to 93.90 bpm in the too-easy group and fell from 99.53 to 97.81 bpm in the too-hard group
Kamikokuryo and colleagues, Sensors 2022	Rehabilitation (hand movement), Oculus Rift S controller	Distance of each movement from its target in an autoencoder latent space	Which of six movements to offer	10 participants without upper extremity pathology (540 movements), plus artificial databases	Sibling Kalman filter exploration had an average cumulative weighted probability error of 7.9, against 19.9 for Boltzmann exploration, with the unsupervised model (8.0 against 20.0 with the semi-supervised model)
Lima and colleagues, BIOSTEC 2022	Cardio-respiratory training in a VR hiking simulator (Levadas from Madeira Island)	Heart rate	Closed-loop adaptation aimed at recommended exertion levels	20 healthy young adults (repeated measures: adaptive VR, non-adaptive VR and non-VR versions)	About 20% more time in the recommended target heart-rate zone, with lower perceived exertion, than the non-adaptive condition; presence, usability and intrinsic motivation were also assessed
de Lima, Silva and Galam, Entertainment Computing 2022	VR horror game	Gameplay data and a machine-learning player model	Use of the horror elements each player fears (fear intensity rather than difficulty)	Not given in the abstract	Fears identified with 79.4% average accuracy for new players; a user study suggests intensified fear and improved immersion and flow
Han and colleagues, COMPSAC 2024	Immersive VR exergame	Not given in the abstract	Game timer	Not given in the abstract	Players in the game with the timer change showed more neutral emotions and improved gaming performance; players without it showed a broader range of detected emotions (sad and happy)
Küntzer and colleagues, VRST 2024 (pilot study)	VR exergame	Heart rate	Difficulty of gameplay challenges	13	Observational results: target heart-rate zones were maintained better than with randomized adjustments, with improved perceived exertion and increased enjoyment
Hein and colleagues, SportsHCI 2025	VR cycling simulator	Heart-rate zones	Visual feedback delivered through non-player characters (not game difficulty)	50 (user study comparing eight visualization designs), 18 (lab study)	Potential to help users stay in their target heart-rate zones; subjective ratings of exertion, enjoyment and motivation remained largely unchanged between conditions
Clocchiatti, Soccini and Inamura, AHs 2026	VR punching game prototype whose outcomes mix the user's motor actions with uncontrollable environmental factors	Not given in the abstract	Difficulty parameters	Not given in the abstract	Preliminary findings: good overall usability and appropriate challenge levels; 50% of participants noticed some changes in the difficulty parameters but attributed them to personal factors
Mantz and colleagues, GMAP 2026 workshop	Two-player VR rhythm exergame	Score difference and mistakes	Rate and strength of bonuses	28 (14 pairs)	Perceived fairness higher with DDA (47.89 against 43.82; one-tailed paired t-test, p = 0.01; Cohen's d = 0.44; the fairness scale was not validated); no significant differences in competence, interest or effort; significantly lower score difference with DDA

Limitations

Andrade and colleagues held that players should not clearly perceive the adaptation. Participants in Mantz and colleagues' focus groups on group VR exergames, by contrast, wanted transparency about DDA during play "to prevent feeling cheated by the game afterward". They feared that the system would overbalance and make the game too easy for the weaker player and too hard for the stronger one, and worried that players could "purposefully play badly in the beginning" to receive an advantage and that less skilled players might feel demotivated if they became aware of a decrease in difficulty. They also rejected automatic decreases in difficulty in favor of player control and preferred subtle changes. The authors cite earlier work finding that DDA "can decrease the self-efficacy of less skilled players and can be perceived as unfair by skilled players", recommend that "DDA Should Provide Opportunities Rather Than Free Bonuses", and conclude that DDA is more accepted in casual games.

Two further studies examined how adjustment affects players' self-perception. In a dance game, Gerling and colleagues found that the explicit balancing approach "commonly used in commercial games" reduced self-esteem and feelings of relatedness in pairs of players, whereas hidden balancing improved self-esteem and reduced the score differential without affecting the game outcome. Constant and Levieux measured confidence with an in-game bet system in three games with logical, motor and sensory difficulty. They compared randomly chosen difficulty with difficulty set by a DDA algorithm and reported that DDA systems "can lead players to high level of overconfidence".

The Li review cautions that performance gains "do not necessarily reflect positive UX outcomes": an adaptive system may reduce task difficulty or constrain user behavior to optimize outcomes while hiding negative experiences such as frustration and anxiety, and questionnaires such as the System Usability Scale, NASA-TLX and the Game Experience Questionnaire are often treated as secondary metrics and reported descriptively rather than integrated into the evaluation of the adaptive mechanism. Individual VR studies report their own limits. Huber and colleagues describe their study as an exploratory proof of concept; because participants spent up to 10 minutes in each maze, the system could take few learning steps, which slowed adaptation, and its user simulation was static. For players already satisfied with the first maze, the results conflicted: ratings of difficulty, exertion and complexity fell for the second maze while mean heart rate rose. Mantz and colleagues say their fairness result "should be seen with caution" because the fairness scale was not validated. Kamikokuryo and colleagues name data as the main limitation of their study: further work had to show whether artificial data can ease the problem of collecting enough data for such deep learning models, and real patient data still had to be tested. They add that the temperature value of the bandit needs further experimentation, and that movements that are too difficult or too easy were not detected or replaced.

The sensors themselves add uncertainty. The Li review cites studies showing that eye-tracking accuracy is influenced by head movement and peripheral viewing conditions, which suggests that many adaptive systems "may rely on signals whose reliability is context-dependent", and that different EEG studies emphasize different frequency bands for similar constructs such as workload, cybersickness and engagement, which makes comparisons across studies difficult. It also observes that fully autonomous proactive systems "require immense computing power and can be unpredictable".

See also

• How to develop VR games with AI
• Procedural generation
• Virtual reality game