Artificial Intelligence approaches for the generation and assessment of believable human-like behaviour in virtual characters

doi:10.1016/j.eswa.2014.05.004

Expert Systems with Applications

Volume 41, Issue 16, 15 November 2014, Pages 7281-7290

https://doi.org/10.1016/j.eswa.2014.05.004 Get rights and content

Highlights

•
Analysis of the believability assessment in video game characters.
•
Differences between First-person and Third-person assessment in Turing test.
•
Cognitive modelling of human-like behaviour generation in video games.
•
Application of Machine Consciousness research to human-like behaviour generation.
•
Results up to 47% of humanness in video games characters.

Abstract

Having artificial agents to autonomously produce human-like behaviour is one of the most ambitious original goals of Artificial Intelligence (AI) and remains an open problem nowadays. The imitation game originally proposed by Turing constitute a very effective method to prove the indistinguishability of an artificial agent. The behaviour of an agent is said to be indistinguishable from that of a human when observers (the so-called judges in the Turing test) cannot tell apart humans and non-human agents. Different environments, testing protocols, scopes and problem domains can be established to develop limited versions or variants of the original Turing test. In this paper we use a specific version of the Turing test, based on the international BotPrize competition, built in a First-Person Shooter video game, where both human players and non-player characters interact in complex virtual environments. Based on our past experience both in the BotPrize competition and other robotics and computer game AI applications we have developed three new more advanced controllers for believable agents: two based on a combination of the CERA–CRANIUM and SOAR cognitive architectures and other based on ADANN, a system for the automatic evolution and adaptation of artificial neural networks. These two new agents have been put to the test jointly with CCBot3, the winner of BotPrize 2010 competition (Arrabales et al., 2012), and have showed a significant improvement in the humanness ratio. Additionally, we have confronted all these bots to both First-person believability assessment (BotPrize original judging protocol) and Third-person believability assessment, demonstrating that the active involvement of the judge has a great impact in the recognition of human-like behaviour.

Introduction

The design and implementation of believable artificial agents, truly indistinguishable from humans, remains an open problem. This challenge has been typically addressed from two interrelated perspectives within cognitive science. On one hand, psychological models of human cognition try to explain how human behaviour is produced. On the other hand, computational models implemented in artificial agents try to replicate to some extent human-like behaviour. In this work, we focus exclusively in the sensorimotor behavioural dimension, setting aside any concerns related to the physical appearance of the artificial agents or their verbal report capabilities.

The imitation game proposed by Turing is the paradigmatic test for believability. However, current state of the art in cognitive and computer sciences has not reached the degree of development in which this test could be considered truly achievable. Therefore, a number of different variations of the original Turing test have been proposed, usually limited Turing tests with relaxed constraints and more specific problem domains. In this paper, we focus in a specific limited version of the Turing test designed for virtual characters and based in a First-Person Shooter (FPS) video game.

From the point of view of the scientific research on human cognition, video game characters are an interesting case of artificial agents because they are easy to implement using the video game industry state of the art tools and their virtual environments can become quite complex, simulating a great variety of contexts and ambient conditions. Furthermore, interaction with real world and with human players is also seamlessly integrated in real-time, as video games are designed to facilitate the prompt interaction between human players and non-player characters (NPC).

While old game character implementations (for instance, Pac-Man ghosts or Space Invaders alien spacecrafts) were based on really simple pre-programmed and scripted behaviours, modern AAA video games are developed to simulate real complex environments and they require engaging, realistic and believable human-like behaviour for their NPCs. Although scripted behaviours might still be acceptable for some specific scenarios, AAA game consumers expect to find synthetic characters at the same level of behavioural realism and unpredictability as evoked by the visual experience of the game.

Generally, human-like behaviour is difficult to both define and test. In fact, the Turing test paradigm stills apply to this problem because no better alternatives have been found to characterise human behaviour. In the realm of computer games, this elusive characterisation might, in principle, be seen easier to define. For instance, human players usually consider disappointing the behaviour of artificial characters for two main reasons (Nareyek, 2004): they are either too intelligent, rational and accurate to be human, or on the contrary, they are too silly. Therefore, the challenge is to find that blurred medium level that characterises human player behaviour.

From the point of view of cognitive science, human-level intelligence and human-like behaviour can be considered as produced by several interrelated psychological processes, ranging from basic activation processes like primary motivations to complex high level cognitive processes such as set shifting and imitation learning. The current knowledge we have about these processes can be used to inspire the design of artificial cognitive architectures. In this paper, we present three different approaches to this sort of inspiration and put them to the test in an adapted version of the Turing test based in a video game (Hingston, 2009). Additionally, we assess the believability (or “humanness”) of these bots using two different assessing methods: First-person and Third-person judges.

The remainder of this paper is structured as follows. In the next section we discuss the problems of assessing believability and describe the testing protocols we have used in this research. In Section 3 we present the different approaches to the design of believable agents, followed in Section 4 by a description of the implementations that we have developed for the believability experiments. Finally, experimental results are presented in Section 5 and discussed in Section 7.

Section snippets

Testing for believability in video games

Testing for human-like behaviour is not straightforward as different observers usually pay attention to different aspects (Arrabales et al., 2012). Therefore the task of judging the believability of a video game character can be approached from the perspective of inter-subjective assessment. In this context there is a key factor to take into account: the possible differences between First-person and Third-person observation. Togelius et al. (2012) argue that believability is better assessed

Different approaches to the design of believable characters

A number of different approaches can be used to address the problem of believable behaviour generation. We can distinguish between two main types of approaches in the design of artificial agent controllers. In one hand, controllers can be built and trained based on data obtained by logs of human behavioural data. On the other hand, controllers can be designed based on models of human cognition. While the former exploit the statistical structure of actions in typical human behaviour, the latter

NPC controller implementations

In this section we describe the three different NPC controllers that we have designed, built and confronted to each other for experimentation.

Experimental results

In the following we summarise the results we have obtained confronting our bot controllers to both the First-person and Third-person believability assessments.

Discussion

As we have shown, three different approaches were designed and compared using two different measurements methods. We can observe that although ANN approach obtains better results in First-person experimentation, in general, cognitive approach CCBotSOAR is the best option. In Fig. 8 we compare the results of the two assessment methods for the three bots. Consistently the First-person method offers a higher humanness ratio for all bots. Applying the Student’s T-test for the paired samples

Conclusions

As introduced in this paper, the automatic generation of human-like behaviour is an enormous challenge, even when addressed in the constrained domain of a video game without verbal interaction. We have explored the application of different control architectures and also new ways to address the problem of assessment. The results obtained clearly indicate that the Third-person approach to assessment is much more demanding in this context. Therefore, we plan to perform more extensive testing using

Acknowledgements

The research reported here has been in part supported by the project TIN2011-24660, funded by the Spanish Ministry of Science and Innovation, and the project FCT-13-7848, funded by the Spanish Foundation for Science and Technology http://human-machine.unizar.es/. This work is also supported by the Spanish MICINN projects TRA2011-29454-C03-01 and TIN2011-25606. The authors also want to thank Philip Hingston for providing the BotPrize competition testing environment.

References (28)

M.L. Anderson
Embodied cognition: A field guide
Artificial Intelligence
(2003)
J.E. Laird et al.
Soar: An architecture for general intelligence
Artificial Intelligence
(1987)
M. Rocha et al.
Evolution of neural networks for classification and regression
Neurocomputing
(2007)
Amato, J., Sappa, A. D., Fornés, A., Lumbreras, F., Lladós, J. (2013). Divide and conquer: atomizing and parallelizing...
Amato, A., Lumbreras, F., Sappa, A. D. (2014) A general-purpose crowdsourcing platform for mobile devices. In...
J.R. Anderson
Rules of the mind
(1993)
M. Arnold
Emotions and Personality
(1960)
R. Arrabales et al.
Consscale fps: Cognitive integration for improved believability in computer game bots
R. Arrabales et al.
A machine consciousness approach to the design of human-like bots
B.J. Baars
In the theatre of consciousness: Global workspace theory, a rigorous scientific theory of consciousness
Journal of Consciousness Studies
(1997)

D. Dennett

Consciousness explained, penguin science

(1993)

P.A. Ertmer et al.

Behaviorism, cognitivism, constructivism: Comparing critical features from an instructional design perspective

Performance Improvement Quarterly

(1993)

J. Gemrot et al.

Pogamut 3 can assist developers in building ai (not only) for their videogame agents

P. Hingston

A turing test for computer game bots, computational intelligence and AI in games

IEEE Transactions on

(2009)

Cited by (32)

Are you for real? Decoding realistic AI-generated faces from neural activity
2022, Vision Research
Citation Excerpt :
Given that observers are already struggling with differentiating between fake and real faces, it is of immediate and practical concern to further investigate the important ways in which the brain can tell the two apart. It is becoming increasingly possible to rapidly and effortlessly generate realistic fake images, videos, writing, and multimedia that are practically indiscernible from real (Radford et al., 2019; Maras & Alexandrou, 2019; Asensio et al., 2014; Ledig et al., 2017). This capacity is only going to become more widespread and has profound implications for cybersecurity, fake news, detection bypass, and social media (Damiani, 2019; Fletcher, 2018; Maddocks, 2020).
Can we trust our eyes? Until recently, we rarely had to question whether what we see is indeed what exists, but this is changing. Artificial neural networks can now generate realistic images that challenge our perception of what is real. This new reality can have significant implications for cybersecurity, counterfeiting, fake news, and border security. We investigated how the human brain encodes and interprets realistic artificially generated images using behaviour and brain imaging. We found that we could reliably decode AI generated faces using people's neural activity. However, while at a group level people performed near chance classifying real and realistic fakes, participants tended to interchange the labels, classifying real faces as realistic fakes and vice versa. Understanding this difference between brain and behavioural responses may be key in determining the 'real' in our new reality. Stimuli, code, and data for this study can be found at https://osf.io/n2z73/.
The evolution of the web and netnography in tourism: A systematic review
2019, Tourism Management Perspectives
Citation Excerpt :
For instance, machines can recognise facial expressions and react to them. Moreover, these technologies also have the facility to add sensorial emotions to avatars for virtual interactions, easing interactions between humans and virtual humans (Llargues Asensio et al., 2014). The evolution of the Web and its application in tourism provides various opportunities for researchers to study the behaviour of consumers and service providers.
Despite recognition of an ever-growing number of Web users, there have been few attempts made to design new empirical material collection methods using various online platforms. Netnography is among the online qualitative methods to have received the least attention from tourism scholars. This systematic review gives an overview of the current gap in netnography studies in tourism that have been published in different disciplines. The findings of this study reveal that the majority of authors conduct their research on Web 2.0. Discussion of the current Web platforms in tourism and possible opportunities for conducting netnography studies are further considered. The contribution of this study is twofold. Firstly, it expands our knowledge of using information technology in tourism to conduct online studies in a general sense. Secondly, it increases our understanding of specific new methods for conducting netnographic studies.
Netnography in tourism – Beyond Web 2.0
2018, Annals of Tourism Research
Citation Excerpt :
For example, facial expressions can be personalized and added to the avatars in real time by using specific headphones. Furthermore, real-time interactions between humans and virtual humans become possible due to the artificial automatic human-like behavior (Llargues Asensio et al., 2014). The virtual human museum guide in the ‘Museum of Science in Boston’ is a tangible example of how visitors can interact with virtual machines (Traum et al., 2012).
Taking the fiction out of science fiction: (Self-aware) robots and what they mean for society, retailers and marketers
2018, Futures
Citation Excerpt :
Since the mid-1950s, AI research has been focused on creating “thinking machines” but developments in the 20th century were slower than anticipated (Baum, Goertzel, & Goertzel, 2011). However, over the last few years there has been a resurgence of research interest on the original goals of AI; that is, creating AI with human-like general intelligence as well as robots that will be hosting this AI (Cassimatis, Mueller, & Winston, 2006; Llargues Asensio et al., 2014; Shrobe, 2014). Experts suggest that AI will be fully self-aware within the next few decades, have a sense of self and be able to engage in self-expression (Baum et al., 2011; Kurzweil, 1999; Lewis et al., 2011).
The development of human-like artificial intelligence has been a goal of computer scientists since the 1950s. Recent technological developments have led to an increased interest in the discourse on human-robot interactions, as robots equipped with artificial intelligence are expected to take a more prominent role in society. This article links insights from psychology and consumer behavior with recent developments in human-robot interactions. In doing so, the article first argues that humanoid robots will have a vast impact on society, specifically on the healthcare, educational and relationship sectors. The article also presents specific implications for retailers and marketers. Moreover, the article outlines examples on how human-robot interactions in the aforementioned sectors may be shaped once humanoid robots with artificial intelligence develop human like self-awareness. Finally, the article offers a viewpoint on how self-aware humanoid robots may even become active members of society that use brands as a means of self-expression; thus, suggesting that the notion of the consumer should be extended from humans to include self-aware humanoid robots.
Training my car to see using virtual worlds
2017, Image and Vision Computing
Citation Excerpt :
For instance, in [134] the computer learns to play Atari games using deep and reinforcement learning. In [135], an artificial agent self-learned how to reproduce human-like behavior while playing a first-person shooter video game (BotPrize competition). In [136], abstract scenarios made with clip art are used to learn unwritten common sense.
Computer vision technologies are at the core of different advanced driver assistance systems (ADAS) and will play a key role in oncoming autonomous vehicles too. One of the main challenges for such technologies is to perceive the driving environment, i.e. to detect and track relevant driving information in a reliable manner (e.g. pedestrians in the vehicle route, free space to drive through). Nowadays it is clear that machine learning techniques are essential for developing such a visual perception for driving. In particular, the standard working pipeline consists of collecting data (i.e. on-board images), manually annotating the data (e.g. drawing bounding boxes around pedestrians), learning a discriminative data representation taking advantage of such annotations (e.g. a deformable part-based model, a deep convolutional neural network), and then assessing the reliability of such representation with the acquired data. In the last two decades most of the research efforts focused on representation learning (first, designing descriptors and learning classifiers; later doing it end-to-end). Hence, collecting data and, especially, annotating it, is essential for learning good representations. While this has been the case from the very beginning, only after the disruptive appearance of deep convolutional neural networks that it became a serious issue due to their data hungry nature. In this context, the problem is that manual data annotation is a tiresome work prone to errors. Accordingly, in the late 00’s we initiated a research line consisting of training visual models using photo-realistic computer graphics, especially focusing on assisted and autonomous driving. In this paper, we summarize such a work and show how it has become a new tendency with increasing acceptance.
Virtual Worlds for Learning in Metaverse: A Narrative Review
2024, Sustainability (Switzerland)

View all citing articles on Scopus

View full text