4) Publication Type: Table XXII displays the number ofpublications according to the publishing channel and machinelearningmethod: 60.98% of the included articles were made available through international conferences, workshops, andsymposia; 19.51% was published as chapters in books; 12.20%appeared in international journals with impact factor; theremaining 9.76% was published in international journals.
V. DISCUSSION AND CONCLUSIONThe number and the distribution of papers identified using thesearch terms listed in this paper confirm the large and increasinginterest in the serious games field related to artificial intelligentalgorithms and techniques. Our inclusion criteria identified129 papers providing evidence for the development and integrationof some specific AI algorithms related to fields of decisionmaking and machine learning.The papers selected for this review were very diverse interms of their area of research and application, market andpurpose, and project and delivery type, reflecting the variedbackgrounds of the researchers and their different interests inserious games. However, all of them have the AI field in common.This, together with the multifeature analysis performedin this paper, helped the authors to create a common analysisframework for organizing and comparing all the includedstudies.The results are quite consistent between the two main sectionsof this paper, which may help to establish some trendsregarding the use of decision-making and machine learningalgorithms in serious games design and development.A. Algorithms UsedConsidering Section III, the number of articles includedunder the umbrella of decision trees is considerably higher thanthe number of studies found for the rest of algorithms. Decisiontrees are computationally undemanding, and have been founduseful as components of intelligent systems [161].However, in Section IV, the included number of articles wasquite balanced between naïve Bayes classifiers, artificial neuralnetworks, and case-based reasoning algorithms. SVM algorithmshad the lower inclusion rate in Section IV. The vastmajority of the SVM studies in this review were focused onclassifying, evaluating, or predicting users’ states and behaviorsusing biological sensors. SVMs are supervised classifiers thathave been improved to successfully work with limited quantityand quality of training samples [162], which may help in theirimplementation with biological data.
B. AI FlagshipSome 65% of articles included in both Sections III and IVbelong to the “PEM—gameplay” flagship. This is an interestingoutcome since both of the reviewed sections produced the sameresult. According to Yannakakis, gameplay-based PEM is theless intrusive and most computationally efficient approach forgames [19], which may be the reason for its high incidence inthis review.“PEM—subjective” flagship did not have representation inthis review, which may be because the authors were morefocused on algorithm-centered studies.C. AI UsageAI techniques were applied with a wide variety of final purposesfor each article. The most common implementations werefor altering the gameflow or for assessing/classifying users’state and behavior while playing. The production of intelligentserious games that dynamically adapt themselves to users’needs and performance have been proved to be efficient in termsof improvement comparisons [163].D. Serious Games MarketEducation and health markets were the most widespreadmarkets considering the reviewed algorithms. No recent keyfigures were found regarding the serious games market analysis.However, the market study published by IDATE whichranges from 1952 to 2009 proved that serious games in mostcases were designed for education [164].E. Serious Games PurposeEdugame-related purposes were the most employed purposein this review. This may be because the use of serious games forlearning purposes is already established [165].F. Platform/DeliveryThe vast majority of the articles included in both decisionmakingand machine learning categories were designed and/ordeveloped for PC. Studies that involved online or mobile seriousgames were in discreet middle distance. Similar resultswere found by Connolly et al. in their review about computerand serious games [165].However, in recent years, the number of smartphones sold toend users worldwide has increased sharply. In 2013, the numberof smartphones sold to customers increased by 50% from2011. This means that almost 20% of world’s population owneda smart device. This figure is expected to grow to 34% by 2017[166]. This increase in the use of portable devices connectedto the Internet could mark a change in trends regarding seriousgames delivery platform.
V. DISCUSSION AND CONCLUSIONThe number and the distribution of papers identified using thesearch terms listed in this paper confirm the large and increasinginterest in the serious games field related to artificial intelligentalgorithms and techniques. Our inclusion criteria identified129 papers providing evidence for the development and integrationof some specific AI algorithms related to fields of decisionmaking and machine learning.The papers selected for this review were very diverse interms of their area of research and application, market andpurpose, and project and delivery type, reflecting the variedbackgrounds of the researchers and their different interests inserious games. However, all of them have the AI field in common.This, together with the multifeature analysis performedin this paper, helped the authors to create a common analysisframework for organizing and comparing all the includedstudies.The results are quite consistent between the two main sectionsof this paper, which may help to establish some trendsregarding the use of decision-making and machine learningalgorithms in serious games design and development.A. Algorithms UsedConsidering Section III, the number of articles includedunder the umbrella of decision trees is considerably higher thanthe number of studies found for the rest of algorithms. Decisiontrees are computationally undemanding, and have been founduseful as components of intelligent systems [161].However, in Section IV, the included number of articles wasquite balanced between naïve Bayes classifiers, artificial neuralnetworks, and case-based reasoning algorithms. SVM algorithmshad the lower inclusion rate in Section IV. The vastmajority of the SVM studies in this review were focused onclassifying, evaluating, or predicting users’ states and behaviorsusing biological sensors. SVMs are supervised classifiers thathave been improved to successfully work with limited quantityand quality of training samples [162], which may help in theirimplementation with biological data.
B. AI FlagshipSome 65% of articles included in both Sections III and IVbelong to the “PEM—gameplay” flagship. This is an interestingoutcome since both of the reviewed sections produced the sameresult. According to Yannakakis, gameplay-based PEM is theless intrusive and most computationally efficient approach forgames [19], which may be the reason for its high incidence inthis review.“PEM—subjective” flagship did not have representation inthis review, which may be because the authors were morefocused on algorithm-centered studies.C. AI UsageAI techniques were applied with a wide variety of final purposesfor each article. The most common implementations werefor altering the gameflow or for assessing/classifying users’state and behavior while playing. The production of intelligentserious games that dynamically adapt themselves to users’needs and performance have been proved to be efficient in termsof improvement comparisons [163].D. Serious Games MarketEducation and health markets were the most widespreadmarkets considering the reviewed algorithms. No recent keyfigures were found regarding the serious games market analysis.However, the market study published by IDATE whichranges from 1952 to 2009 proved that serious games in mostcases were designed for education [164].E. Serious Games PurposeEdugame-related purposes were the most employed purposein this review. This may be because the use of serious games forlearning purposes is already established [165].F. Platform/DeliveryThe vast majority of the articles included in both decisionmakingand machine learning categories were designed and/ordeveloped for PC. Studies that involved online or mobile seriousgames were in discreet middle distance. Similar resultswere found by Connolly et al. in their review about computerand serious games [165].However, in recent years, the number of smartphones sold toend users worldwide has increased sharply. In 2013, the numberof smartphones sold to customers increased by 50% from2011. This means that almost 20% of world’s population owneda smart device. This figure is expected to grow to 34% by 2017[166]. This increase in the use of portable devices connectedto the Internet could mark a change in trends regarding seriousgames delivery platform.