Материал: искусственный интеллект

120 R. A. F. Cunha et al.

Table 2. Doubt emotion truth table

5 Quantum Wheel of Emotions

The concept of “wheel of emotions” introduced by Plutchik et al. [7] pictures the idea that a complex emotional state is the composition of elementary emotions. This picture can be interpreted in a quantum-like way using the quantum state vector |ψ . Each qubit (2-dimensional) quantum state can be mapped to a point on the surface of the Bloch unit sphere:

where φ and θ are the spherical angles. In order to simplify interpretation, the coe cients (associated with the degree of truth) multiplying the base states are taken as real numbers. The points of the vector are thus placed on a circle corresponding to a quantum wheel of emotions.

We summarized our simulation results by associating the di erent observed behaviors to a sector in the wheel as shown in Fig. 4. Other emotions not presented in Fig. 3 have been simulated such as for example: Interest, Curiosity, Distraction, Fear, Worship and Sadness. The two latter ones have been obtained using a circuit that combines the standard quantum gates H and CNOT.

The quantum wheel of emotions thus allows a continuous set of emotional states. A small perturbation in the angle of the input state |ψ due to environmental factors, even if still inside the same emotional sector, will correspond to small changes in the vehicle’s behavior. The measurement of the input state implies the collapse of |ψ to a specific point of the wheel, and thus we can say that, in this aspect, the vehicle behaves as a quantum-like system. Furthermore, the fuzzy aspect of the system arouses naturally since the collapse can involve any state belonging to the continuous surface of the wheel. These observations can be compared to the observed similarities between neural network models and quantum systems. In particular, it has been suggested that it is possible to implement quantum learning algorithms dedicated to fuzzy qubits [8] where the weighted sums of inputs of a neuron correspond to the superposition of quantum states at the input of a quantum circuit and the quantum wave function collapse corresponds to the threshold activation of a neuron.

Fig. 4. Quantum Wheel of Emotions. For a given emotion, we associate quantum operators controlling the speed of the left and the right wheel respectively. (Anger)

FB , FA ; (Passion) FB = A , FA = B ; (Love) F ¯, F ¯ ; (Interest) F ¯ , F ¯; (Curios-

A B B A

ity) FA = B , FB = A ; (Distraction)FA = B , FA XOR B; (Apprehension) FB = A , FA XOR B; (Worship) H H , FB ; (Sadness) CNOT, CNOT; (Fear) FA , FB .

6 Discussion and Conclusion

The purpose of this research is to show the multiplicity of behaviors obtained by using fuzzy logic along with quantum logical gates in the control of simple Braitenberg Vehicle agents. The number of cases becomes intractable in simple theoretical approaches with increasing complexity. A computer simulation is mandatory and allows us to abstract the complexity by observing the motion of the vehicles and use it for illustrative purposes. At the same time, we see that by changing and combining di erent quantum control gates we can tune small changes in the vehicle’s behavior, and hence get specific features around the main basic BV emotions of Fear, Aggression, Love and Explore. By tweaking these quantum gates, one can also obtain a vehicle that has a mixture of multiple emotions.

Further extensions to this project can be imagined. Currently, when the vehicles collide, their respective control operators could change in order to reflect a quantum-like entanglement behavior due to interaction. It would be interesting to entangle the vehicles so that the behavior of one vehicle depends upon the current state of the environment from the perspective of other vehicles even after they separate after the collision. It could also be interesting to explore the Braitenberg vehicles using di erent types of stimuli (instead of only light) and sensors. Also a formalization of the quantum BV components as quantum neural networks could lead to new investigation strategies and could benefit researches in machine learning algorithms related to emotion analysis.

122 R. A. F. Cunha et al.

7 Credits and Acknowledgements

The first coauthor is under the Brafitec scholarship, CAPES Foundation, Ministry of Education of Brazil, Brasilia 70.040-020, Brazil.

We want to thank the AFSCET (Association Fran¸caise de Science des Syst`emes) for permitting us to present the idea of this work at the WOSC (World Organisation of Systems and Cybernetics) 2017 Congress in Rome Italy.

We are grateful to Francesco Galofaro of LUB for fruitful discussions on logic and semantics and their link to quantum theory and for pointing out that the late Professor Valentin von Braitenberg was one of the academic founding members of LUB (Libera Universit`a di Bolzano/Freie Universit¨at Bozen, Italy).

References

1.Braitenberg, V.: Vehicles - Experiments in Synthetic Psychology. MIT Press, Cambridge (1986)

2.Benio , P.: Quantum robots and environments as a first step towards a quantum mechanical description of systems that are aware of their environment and make decisions. Phys. Rev. A 58(2), 893–904 (1998)

3.Raghuvanshi, A., Fan, Y., Woyke, M., Perkowski, M.: Quantum robots for teenagers. In: 37th International Symposium on Multiple-Valued Logic ISMVL (2007)

4.Dubois, F., To ano, Z.: Eigenlogic: a quantum view for multiple-valued and fuzzy systems. In: de Barros, J.A., Coecke, B., Pothos, E. (eds.) QI 2016. LNCS, vol. 10106, pp. 239–251. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-52289-0 19

5.Kagan, E., Rybalov, A., Sela, A., Siegelmann, H.: Probabilistic control and swarm dynamics in mobile robots and ants. In: Biologically-Inspired Techniques for Knowledge Discovery and Data Mining, January, pp. 11–13 (2014)

6.To ano, Z., Dubois, F.: Interpolating binary and multivalued logical quantum gates. In: Proceedings of the 4th International Electronic Conference on Entropy and Its Applications. MDPI Proceedings, vol. 2, no. 4, p. 152 (2018)

7.Plutchik, R.: The nature of emotions. Am. Sci. 89(4), 334–350 (2001)

8.Hannachi, M.S., Dong, F., Hirota, K.: Emulating quantum interference and quantum associative memory using fuzzy qubits. In: Proceedings of the IEEE International Conference on Computational Cybernetics ICCC 2007, pp. 39–45 (2007)

Moral Dilemmas for Artificial Intelligence: A Position Paper on an Application

of Compositional Quantum Cognition

Camilo M. Signorelli1,2,5(&) and Xerxes D. Arsiwalla3,4,5

1 Department of Computer Science, University of Oxford, Oxford, UK

camiguel@uc.cl

2 Cognitive Neuroimaging Lab, INSERM U992, NeuroSpin,

Gif-sur-Yvette, France

3Barcelona Institute of Science and Technology, Barcelona, Spain 4 Institute for Bioengineering of Catalonia, Barcelona, Spain

5 Pompeu Fabra University, Barcelona, Spain

Abstract. Traditionally, the way one evaluates the performance of an Artificial Intelligence (AI) system is via a comparison to human performance in specific tasks, treating humans as a reference for high-level cognition. However, these comparisons leave out important features of human intelligence: the capability to transfer knowledge and take complex decisions based on emotional and rational reasoning. These decisions are influenced by current inferences as well as prior experiences, making the decision process strongly subjective and “apparently” biased. In this context, a definition of compositional intelligence is necessary to incorporate these features in future AI tests. Here, a concrete implementation of this will be suggested, using recent developments in quantum cognition, natural language and compositional meaning of sentences, thanks to categorical compositional models of meaning.

Keywords: Moral dilemmas Moral test Turing test Artificial Intelligence Compositional semantics Natural language Quantum cognition

1 Introduction

Moral dilemmas and a general intelligence definition have been recently suggested as an alternative to current AI tests [1]. Usually, Intelligence is interpreted regarding particular and efficient behaviours which can be measured in terms of performing or not these behaviours. One example is the Turing test [2], in fact, the first approach grounded on human behaviour and the most well-known and controversial test for AI. Other examples are challenging machines in games like chess or Go [3], and testing AI programs with dilemmas [4], theoretically, demanding a more complex level of information processing. Nevertheless, all these approaches lack a general definition of intelligence as a minimal requirement to measure intelligence [5–7], and restrict AI only to humankind “intelligence” without including key features of a true human intelligence. Therefore, this position paper is going to shortly introduce a new strategy

© Springer Nature Switzerland AG 2019

B. Coecke and A. Lambert-Mogiliansky (Eds.): QI 2018, LNCS 11690, pp. 123–138, 2019. https://doi.org/10.1007/978-3-030-35895-2_9

124 C. M. Signorelli and X. D. Arsiwalla

and research program to quantify a probably more complete and inclusive evaluation for future advances in AI.

The discussion will start with a preliminary definition of general intelligence; some key ingredients in human intelligence will be analyzed to finally arrive at more general principles of compositional intelligence. Then, a Moral test emerges naturally, as an option to effectively identify a combination of different and important features in human intelligence. Additionally, the issue about how to implement a Moral test is sketched, using new advances in the categorical compositional model of meaning (CCMM) in natural language [8] and the emergent field of Quantum Cognition (QC), associated with cognition and decision making under uncertainty [9].

2 Towards a General Definition of Intelligence

2.1What Is Intelligence?

A useful way to approach this question is by conceptualizing general intelligence as the capability of any system to take advantage of its environment in order to achieve a specific or general goal [1]. Biologically speaking this goal would be surviving, or in reductive terms: trying to keep the autonomy and potential reproduction of the system; while the goal in machines can be solving a specific task or problem using internal and external resources. This advantage would take place as a balance of these resources or in other words, reducing disequilibrium between them. This balance is managed internally, as the cognitive architecture/process to deal with internal and external demands, using both internal and external feedback. If any animal, human or machine break the balance between internal and external resources, as for example achieving their particular needs at the expense of the total annihilation of their environment and resources, this animal, human or machine is leading its own annihilation, which is not particularly intelligent. Therefore, a general intelligence is defined here as the balance (reducing disequilibrium) between these external and internal resources like an embodied system [10]. This general definition can incorporate living beings as well as robots and computers, and in this way, intelligence is general enough to include different kind of intelligence, multiple intelligence, contextual influences and different kind of systems with different degrees of intelligence [1]. Balance, as a “relative efficiency”, is considered as part of a general feature of intelligent systems and it can be linked with a physical view of intelligence using entropy or complexity [6, 7, 11, 12]. The mathematical description of these concepts is expected in future developments.

2.2Human Intelligence and Quantum Cognition

A preliminary definition of human intelligence can be suggested as the ability to benefit or gain advantage from their social environment while maintaining autonomy [1, 13]. This requires a balance or equilibrium between rational and emotional information processes [1, 10]. Hence, humans would solve problems by trying to incorporate different types of information and balancing both internal and external resources.