A wide range of specific tasks can be handled efficiently by present day robots and machine learning algorithm. Playing a game of Go is a typical example, with CT-scan classification  and micro-helicopter piloting  being two other tasks that are accomplished with comparative modest computational resources. As a next step one may consider agents that have not just a one, but a range of capabilities. like being able to play several distinct board games. Classically, as in most today’s applications, a human operator interacting with the program decides which of the agents functionality he or she wants to access. Humanized agents may however be envisioned to function without direct supervision, also with respect to task selection . Autonomous agents that are endorsed with the capability to connect on their own to either a Go or a chess server could be matched f.i. with either a human opponent, upon entering the queuing system, or with another board-playing program. The problem is then how to allocate time , namely how to decide which type of game to play.
Deciding what to do is a cornerstone of human activities, which implies that frameworks for multi-task situations deserve attention. A possible route is to define and to maximize an overarching objective function, which could be, for example, to play alternatively Go or chess in order to improve the respective levels of expertise, as measured, f.i., by the respective win rates. This example suggest that time-allocation frameworks need two components, see Fig. 1:
A set of criteria characterizing tasks that have been executed, the experience of the agent.
A set of rules for task switching that are based fully or in part on experiences.
Implementations may distinguish further between dedicated and humanized settings. For a specific dedicated application appropriate hand-picked sets of evaluation criteria and switching rules can be selected. Within this approach the resulting overall behavior can be predicted and controlled to a fair extend. Being hand-crafted, the disadvantage is that extensions and transfer to other domains demand in general extensive reworks.
Consider an agent with two initial abilities, to play Go and chess via internet servers. The respective winning rates could be taken in this case as appropriate evaluation criteria, other may be a challenge (close games) and boredom (games lasting forever). As an extension, the agent is provided with a connection to a chat room, where the task is to answer questions. Humans would sent in chess board positions and the program provide in return the appropriate analysis, e.g. in terms of possible moves and winning probabilities. The program has then three possibilities, as shown in Fig.1, to play Go or chess, and to connect to the chat room, with the third task differing qualitatively from the first two. Time allocation frameworks designed specifically for the first two options, to play Go or chess, would most probably cease to work when the chat room is added, f.i. because winning ratios are not suitable for characterizing a chat session. Here we argue that a characteristic trait of humanized computing is universality, which translates in the context of time allocation frameworks to the demand that extensions to new domains should be a minor effort.
Of particular interest to humanized time allocation frameworks is emotional control, which is known to guide human decision making. Starting with an overview, we will discuss first the computational and neurobiological role of mammalian emotions, stressing that algorithmic implementations need to reproduce functionalities and not qualia like fear and joy. A concrete implementation based on the stationarity principle is then presented in a second step. Synthetic emotions correspond in this framework to a combination of abstract evaluation criteria and motivational drives that are derived from the objective to achieve a predefined time-averaged distribution of emotional activities, the ‘character’ of the agent.
An alternative to the here explored route to multi-task problems is multi-objective optimization , a setting in which distinct objectives dispose of individual utility function that need to be optimized while respecting overall resource limitation, like the availability of time. We focus here on emotional control schemes, noting that emotional control and multi-objective optimization are not mutually exclusive.
Ii Computational role of emotions
Emotions have emerged in the last decades as indispensable preconditions for higher cognition [7, 8], with the reason being that the core task of emotional response is not direct causation of the type “fleeing when afraid”, akin to a reflex, but the induction of anticipation, reflection and cognitive feedback . In general, being afraid will not result in a direct behavioral response, but in the allocation of cognitive resources to the danger at hand.
The interrelation between emotion and cognition is two-faced. Emotions prime cognitive processes , being controlled in return by cognition . The latter capability, to regulate emotions , f.i. when restraining one’s desire for unhealthy food, is so pronounced that it can be regarded to be a defining characteristics our species . With regard to synthetic emotions, it is important to note that the cogno-emotional feedback loop present in our brain implies that emotional imprints are induced whenever cognitive capabilities are used to pursue a given goal, such as playing and winning a game of Go .
On a neuronal level one may argue, that the classical distinction between affective and cognitive brain regions is misleading . Behavior should be viewed instead as a complex cogno-emotional process that is based on dynamic coalitions of brain areas , and not on the activation of a specific structure, such as the amygdala . This statement holds for the neural representations of the cognitive activity patters regulating emotional reactions, which are not localized in specific areas, but distributed within temporal, lateral frontal and parietal regions .
The mutual interrelation of cognitive and emotional brain states suggests a corresponding dual basis for decision making . Alternative choices are analyzed using logical reasoning, with the outcome being encoded affectively . Here we use ‘evaluation criteria’ as a generic term for the associated emotional values. Risk weighting has similarly both cognitive and emotional components , where the latter are of particular importance for long-term, viz strategic decision taking . One feels reassured if a specific outlook is both positive and certain, and uncomfortable otherwise.
The picture emerging from affective neuroscience studies is that the brain uses deductive reasoning for the analysis of behavioral options and emotional states for the respective weighting. A larger number of distinct types of emotional states , like anger, pride, fear, trust, etc, is consequently needed when the space of accessible behavioral options increases .
Iii Cogno-emotional architectures
A minimal precondition for application scenaria incorporating a basic cogno-emotional feedback loop is the option for the program to switch between tasks . An example is a multi-gaming environment for which the program decides on its own, as detailed out further below, which game to play next.
Iii-a Multi-gaming environments
We consider an architecture able to play several games, such as Go, chess, Starcraft or console games like Atari. The opponents may be either human players that are drawn from a standard internet-based matchmaking systems, standalone competing algorithms or agents participating in a multi-agent challenge setup . Of minor relevance is the expertise level of the architecture and whether game-specific algorithms are used. A single generic algorithm , such as standard Monte Carlo tree search supplemented by a value and policy generating deep network , would do the job. For our purpose, a key issue is the question whether the process determining which game to play is universal, in the sense that it can be easily adapted when the palette of tasks is enlarged, f.i. when the option to connect to a chat room is added.
For a complete cogno-emotional feedback loop an agent able to reason logically on an at least rudimentary level would be needed. This does not hold for the application scenario considered here. As a consequence, one may incorporate the feedback of the actions of the agent onto its emotional states and the emotional priming of the decision process, but not a full-fledged cognitive control of emotions.
Iii-B Emotional evaluation criteria
In a first step one has to define the qualia of the emotional states and how they are evaluated, viz the relation of distinct emotions to experiences. The following definitions are examples.
Satisfaction. Winning a game raises the satisfaction level. This could hold in particular for complex games, that is for games that are characterized, f.i., by an elevated diversity of game situations.
Challenge. Certain game statistics may characterize a game as challenging. An example would be games for which the probability to win dropped temporarily precariously low.
Games for which the probability to win remains constantly high could be classified as boring or, alternatively, as relaxing. The same holds for overly long games.
Emotions correspond to value-encoding variables, denoted here with , and , for satisfaction, challenge and boredom. Games played are evaluated using a set of explicit evaluation criteria, as formulated above. An important note is that the aim of our framework is to model key functional aspects of human emotions, which implies that there is no need, as a matter of principle, for the evaluation criteria to resemble human emotions in terms of their qualia. The latter is however likely to make it easier to develop an intuitive understanding of emotionally-driven robotic behavior.
Iii-C Direct emotional drivings vs. emotional priming
Standard approaches to modeling synthetic approaches often assume that emotional state variables are explicit drivers of actions , either directly or via a set of internal motivations . Here we are interested in contrast in frameworks that are generic in the sense that behavior is only indirectly influenced by emotional states .
In our case the agent updates in a first step its experience. For every type of activity, say when playing Go, the probability that a game of this type is challenging, boring or satisfying is continuously updated. It could be, e.g., that Go games are typically more challenging and less boring than chess games. Based on this set of data, the experience, the next game will be selected with the aim to align experience as close as possible with the ‘character’ of the agent, as defined in the following.
Iii-D Aligning experience with character
We define the character of the agent as a probability distribution of emotional states,
where are the target frequencies to experience a given emotional state. Agents with a large would prefer for example challenging situations. The overall objective function of the agent is to align experience with his character.
On a basic level, experience is expressed as a set of probability distribution functions,
where is the number of possible activities (playing Go, chess, connecting to a chat room, …). For every option the agent records, as illustrated in Fig. 2, the probability for the activity to be satisfying/challenging/boring (). Defining with the likelihood to engage in activity , the overall experience is given as
where the are defined in (2). The global objective, to align character and experience
, is achieved by minimizing the Kullback-Leibler divergence betweenand with respect to the . This strategy, which corresponds to a greedy approach, can be supplemented by an explorative component that allows to sample new opportunities . Modulo exploration, an activity is hence selected with probability
Iii-E Stationarity principle
Our framework is based on aligning two probability distribution functions, and , an information-theoretical postulate that has been denoted the ‘stationarity principle’  in the context of neuronal learning . It states, that not the activity as such should be optimized, but the distribution of activities. The resulting state is then varying in time, but stationary with respect to its statistical properties. The underlying principle of the here presented framework corresponds to ‘time allocation via emotional stationarity’ (TAES). Within this approach the character of the agent serves as a guiding functional, a stochastic implementation of the principle of guided self-organization .
Iii-F Motivational drives
Up to now we considered purely stochastic decision making, namely that activities are selected probabilitistically, as determined by the selection probabilities . An interesting extension are deterministic components that correspond to emotional drives. Considering finite time spans, we denote with the relative number of times that emotion has been experienced over the course of the last activities. Ideally, the trailing averages converge to the desired frequencies . Substantial fluctuations may however occur, for example when the agent is matched repeatedly to opponents with low levels of expertise, which may lead to an extended series of boring games. The resulting temporary discrepancy,
between desired and trailing emotion probabilities can then be regarded as an emotional drive. Stochastically, averages out, as far as possible, when selecting appropriate probabilities to select an activity . On a shorter time scale one may endorse the agent with the option to reduce excessive values of by direct action, viz by selecting an activity characterized by large/small when is strongly positive/negative. This is however only meaningful if the distribution is peaked and not flat. Emotional drives correspond in this context to a additional route for reaching the overall goal, the alignment of experience with character.
Iii-G Including utility maximization
In addition to having emotional motivations, agents will in general be expected to maximize one or more reward functions, like gaining credits for wining games or answering questions in a chat room. Without emotional constraints, the program would just select the most advantageous option, given that all options have already been explored in sufficient depth, in analogy to the multi-armed bandit problem . An interesting constellation arises when rewards are weighted emotionally, e.g. with the help of the Kullback-Leibler divergence between the character and the emotional experience of a given behavioral option ,
Credits received from behavioral options that conform with the character of the agent, having a small , would be given a higher weight than credits gained when engaging in activities characterized by a large . There are then two conflicting goals, to maximize the weighted utility and to align experience with character, for which a suitable prioritization or Pareto optimality may be established .
Instead of treating utility as a separate feature, one may introduce a new emotional trait, the desire to receive rewards, and subsume utility under emotional optimization. Depending on the target frequency to generate utility, the agent will select its actions such that the full emotional spectrum is taken into account. A relative weighting of utility gains, as expressed by (5), is then not necessary.
Computational models of emotions have focused traditionally on the interconnection between emotional stimuli, synthetic emotions and emotional responses . A typical goal is to generate believable behaviors of autonomous social agents , in particular in connection with psychological theories of emotions, involving f.i. appraisal, dimensional aspects or hierarchical structures . Closer to the scope of the present investigation are proposals that relate emotions to learning and such to behavioral choices 
. One possibility is to use homeostatic state variables, encoding f.i. ‘well-being’, for the regulation of reinforcement learning. Other state variables could be derived from utility optimization, like water and energy uptake, or appraisal concepts , with the latter being examples for the abstract evaluation criteria used in the TAES framework. One route to measure well-being consist in grounding it on the relation between short- and long-term trailing reward rates . Well-being can then be used to modulate dynamically the balance between exploitation (when doing well) and exploration (when things are not as they used to be). Alternatively, emotional states may impact the policy .
Going beyond the main trust of research in synthetic emotions, to facilitate human-computer interaction and and to use emotions to improve the performance of machine learning algorithms that are applied to dynamic landscapes, the question that has been asked here regards how an ever ongoing sequence of distinct tasks can be generated by optimizing emotional experience, in addition to reward. Formulated as a time allocation problem, the rational of this approach is drawn mainly from affective neuroscience , and only to a lesser extend from psychological conceptualizations of human emotional responses. Within this setting, the TAES framework captures the notion that a central role of emotions is to serve as abstract evaluation tools that are to be optimized as a set, and not individually. This premise does not rule out alternative emotional functionalities.
Frameworks for synthetic emotions are especially powerful and functionally close to human emotions if they can be extended with ease along two directions. First, when the protocol for the inclusion of new behavioral options is applicable to a wide range of activity classes. This is the case when emotions do not correspond to specific features, but to abstract evaluation criteria. A given activity could then be evaluated as being boring, challenging, risky, demanding, easy, and so on. It is also desirable that the framework allows for the straightforward inclusion of new traits of emotions, such as frustration.
Two agents equipped with the identical framework can be expected to be able to show distinct behaviors, in analogy to the observation that human decision making is generically dependent on the character of the acting person. For synthetic emotions this implies that there should exist a restricted set of parameters controlling the balancing of emotional states in terms of a preferred distribution, the functional equivalent of character. As realized by the TAES framework, the overarching objective is to adjust the relative frequencies to engage in a specific task, such that the statistics of the experienced emotional states aligns with the character.
Choosing between competing reward options can be done using a variety of strategies . An example is the multi-armed bandits problem, for which distinct behavioral options yield different rewards that are initially not known . Human life is characterized in comparison by behavioral options, to study, to visit a friend, to take a swim in the pool, and so on, that have strongly varying properties that come with multi-variate reward dimensions. As a consequence we proposed to define utility optimization not in terms of money-like credits, as it is the case for the multi-armed bandits problem, but on an abstract level. For this one needs evaluation criteria that are functionally equivalent to emotions. In this perspective, life-long success depends not only on the algorithmic capability to handle specific tasks, but also on the character of the agent.
-  D. Silver, et al., “Mastering the game of go without human knowledge,” Nature, vol. 550, no. 7676, pp. 354–359, 2017.
X. W. Gao, R. Hui, and Z. Tian, “Classification of ct brain images based on deep learning networks,”Computer methods and programs in biomedicine, vol. 138, pp. 49–56, 2017.
-  V. Kumar and N. Michael, “Opportunities and challenges with autonomous micro aerial vehicles,” in Robotics Research. Springer, 2017, pp. 41–58.
-  M. Malfaz, Á. Castro-González, R. Barber, and M. A. Salichs, “A biologically inspired architecture for an autonomous and social robot,” IEEE Transactions on Autonomous Mental Development, vol. 3, no. 3, pp. 232–246, 2011.
C. Gros, “Emotional control–conditio sine qua non for advanced artificial intelligences?” inPhilosophy and Theory of Artificial Intelligence. Springer, 2013, pp. 187–198.
-  K. Deb, “Multi-objective optimization,” in Search methodologies. Springer, 2014, pp. 403–449.
-  J. Panksepp, Affective neuroscience: The foundations of human and animal emotions. Oxford university press, 2004.
-  C. Gros, “Cognition and emotion: perspectives of a closing gap,” Cognitive Computation, vol. 2, no. 2, pp. 78–85, 2010.
-  R. F. Baumeister, K. D. Vohs, C. Nathan DeWall, and L. Zhang, “How emotion shapes behavior: Feedback, anticipation, and reflection, rather than direct causation,” Personality and social psychology review, vol. 11, no. 2, pp. 167–203, 2007.
-  J. A. Beeler, R. Cools, M. Luciana, S. B. Ostlund, and G. Petzinger, “A kinder, gentler dopamine… highlighting dopamine’s role in behavioral flexibility,” Frontiers in neuroscience, vol. 8, 2014.
-  K. N. Ochsner and J. J. Gross, “The cognitive control of emotion,” Trends in cognitive sciences, vol. 9, no. 5, pp. 242–249, 2005.
-  M. Inzlicht, B. D. Bartholow, and J. B. Hirsh, “Emotional foundations of cognitive control,” Trends in cognitive sciences, vol. 19, no. 3, pp. 126–132, 2015.
-  D. Cutuli, “Cognitive reappraisal and expressive suppression strategies role in the emotion regulation: an overview on their modulatory effects and neural correlates,” Frontiers in Systems Neuroscience, vol. 8, 2014.
-  M. Miller and A. Clark, “Happily entangled: prediction, emotion, and the embodied mind,” Synthese, vol. 195, no. 6, pp. 2559–2575, 2018.
-  L. Pessoa, “On the relationship between emotion and cognition,” Nature reviews neuroscience, vol. 9, no. 2, p. 148, 2008.
-  ——, “Embracing integration and complexity: placing emotion within a science of brain and behaviour,” Cognition and Emotion, vol. 33, no. 1, pp. 55–60, 2019.
-  ——, “Understanding emotion with brain networks,” Current opinion in behavioral sciences, vol. 19, pp. 19–25, 2018.
-  E. A. Phelps, “Emotion and cognition: insights from studies of the human amygdala,” Annu. Rev. Psychol., vol. 57, pp. 27–53, 2006.
-  C. Morawetz, S. Bode, J. Baudewig, A. M. Jacobs, and H. R. Heekeren, “Neural representation of emotion regulation goals,” Human brain mapping, vol. 37, no. 2, pp. 600–620, 2016.
-  J. S. Lerner, Y. Li, P. Valdesolo, and K. S. Kassam, “Emotion and decision making,” Annual review of psychology, vol. 66, pp. 799–823, 2015.
-  M. Reimann and A. Bechara, “The somatic marker framework as a neurological theory of decision-making: Review, conceptual comparisons, and future neuroeconomics research,” Journal of Economic Psychology, vol. 31, no. 5, pp. 767–776, 2010.
-  A. Panno, M. Lauriola, and B. Figner, “Emotion regulation and risk taking: Predicting risky choice in deliberative decision making,” Cognition & emotion, vol. 27, no. 2, pp. 326–334, 2013.
-  R. Gilkey, R. Caceda, and C. Kilts, “When emotional reasoning trumps iq,” harvard business review, vol. 88, no. 9, p. 27, 2010.
-  H.-R. Pfister and G. Böhm, “The multiplicity of emotions: A framework of emotional functions in decision making,” Judgment and decision making, vol. 3, no. 1, p. 5, 2008.
-  T. Schlösser, D. Dunning, and D. Fetchenhauer, “What a feeling: the role of immediate and anticipated emotions in risky decisions,” Journal of Behavioral Decision Making, vol. 26, no. 1, pp. 13–30, 2013.
-  T. Rumbell, J. Barnden, S. Denham, and T. Wennekers, “Emotions in autonomous agents: comparative analysis of mechanisms and functions,” Autonomous Agents and Multi-Agent Systems, vol. 25, no. 1, pp. 1–45, 2012.
-  M. Samvelyan, T. Rashid, C. S. de Witt, G. Farquhar, N. Nardelli, T. G. Rudner, C.-M. Hung, P. H. Torr, J. Foerster, and S. Whiteson, “The starcraft multi-agent challenge,” arXiv preprint arXiv:1902.04043, 2019.
-  D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel et al., “Mastering chess and shogi by self-play with a general reinforcement learning algorithm,” arXiv preprint arXiv:1712.01815, 2017.
-  L.-F. Rodríguez and F. Ramos, “Computational models of emotions for autonomous agents: major challenges,” Artificial Intelligence Review, vol. 43, no. 3, pp. 437–465, 2015.
-  J. Velsquez, “Modeling emotions and other motivations in synthetic agents,” in Proc. 14th Nat. Conf. Artif. Intell, 1997, pp. 10–15.
-  P. Auer, “Using confidence bounds for exploitation-exploration trade-offs,” Journal of Machine Learning Research, vol. 3, no. Nov, pp. 397–422, 2002.
R. Echeveste, S. Eckmann, and C. Gros, “The fisher information as a neural guiding principle for independent component analysis,”Entropy, vol. 17, no. 6, pp. 3838–3856, 2015.
P. Trapp, R. Echeveste, and C. Gros, “Ei balance emerges naturally from continuous hebbian learning in autonomous neural networks,”Scientific reports, vol. 8, no. 1, p. 8939, 2018.
-  C. Gros, “Generating functionals for guided self-organization,” Guided Self-Organization: Inception, pp. 53–66, 2014.
-  J. Vermorel and M. Mohri, “Multi-armed bandit algorithms and empirical evaluation,” in European conference on machine learning. Springer, 2005, pp. 437–448.
-  C. Gros, Complex and adaptive dynamical systems: A primer. Springer, 2015.
-  O. Sener and V. Koltun, “Multi-task learning as multi-objective optimization,” in Advances in Neural Information Processing Systems, 2018, pp. 527–538.
-  K. R. Scherer, “Emotions are emergent processes: they require a dynamic computational architecture,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 364, no. 1535, pp. 3459–3474, 2009.
-  S. C. Gadanho, “Learning behavior-selection by emotions and cognition in a multi-goal robot task,” Journal of Machine Learning Research, vol. 4, no. Jul, pp. 385–412, 2003.
-  T. M. Moerland, J. Broekens, and C. M. Jonker, “Emotion in reinforcement learning agents and robots: a survey,” Machine Learning, vol. 107, no. 2, pp. 443–480, 2018.
-  J. Broekens, W. A. Kosters, and F. J. Verbeek, “On affect and self-adaptation: Potential benefits of valence-controlled action-selection,” in International Work-Conference on the Interplay Between Natural and Artificial Computation. Springer, 2007, pp. 357–366.
-  T. Kuremoto, T. Tsurusaki, K. Kobayashi, S. Mabu, and M. Obayashi, “An improved reinforcement learning system using affective factors,” Robotics, vol. 2, no. 3, pp. 149–164, 2013.
-  C. Gros, “Emotions, diffusive emotional control and the motivational problem for autonomous cognitive systems,” in Handbook of Research on Synthetic Emotions and Sociable Robotics: New Applications in Affective Computing and Artificial Intelligence. IGI Global, 2009, pp. 119–132.
-  M. I. Jordan and T. M. Mitchell, “Machine learning: Trends, perspectives, and prospects,” Science, vol. 349, no. 6245, pp. 255–260, 2015.