Abstract
The import of computational learning theories and techniques on the ethics of human-robot interaction is explored in the context of recent developments of personal robotics. An epistemological reflection enables one to isolate a variety of background hypotheses that are needed to achieve successful learning from experience in autonomous personal robots. The conjectural character of these background hypotheses brings out theoretical and practical limitations in our ability to predict and control the behaviour of learning robots in their interactions with humans. Responsibility ascription problems, which concern damages caused by learning robot actions, are analyzed in the light of these epistemic limitations. Finally, a broad framework is outlined for ethically motivated scientific inquiries, which aim at improving our capability to understand, anticipate, and selectively cope with harmful errors by learning robots.
Similar content being viewed by others
Notes
Failures of this safety policy in industrial environments are witnessed by a significant number of accidents involving robots in factories and plants. Useful information about robots and safety of human beings is provided in the 8 Jun 2006 issue of The Economist-Technology Quarterly. An electronic copy of the article is available on line at: http://www.economist.com/displaystory.cfm?story_id = 7001829.
The overall rule governing the behaviour of an unsociable robot is relatively easy to state, but its actual implementation raises non-trivial theoretical and technological problems, which include the need for real-time reactivity and motion planning in high-dimensional configuration spaces. Furthermore, the reliability of the proposed solutions usually declines sharply when the environment becomes more and more cluttered, dense, and complex. A survey of effective methods and solutions to such problems can be found in Minguez and Montano (2004); Brock and Khatib (2002); and Kohout (2000).
Several proposals for relaxing some of these constraints have been advanced, but the modified learning algorithms are usually intractable for all but the smallest problems. Even though a few algorithms (see, e.g., Pineau and Gordon 2005) solved some such computational problems and demonstrated competitive performances in limited tasks, their use is still far from being widespread in robotics.
Background conjectural assumptions about the environment play a crucial role here too, insofar as successful application of RL learning depends on the correctness of these assumptions about the environment. A detailed argument to this effect is provided, in connection with RL algorithms for adapting navigation control strategies in behaviour-based robotic architectures, in Datteri et al. (2006).
A solution for a class of learning problems is the specification of a learning algorithm that can be trained by means of a suitable set of training examples.
From the standpoint of computational complexity theory this is usually taken to mean that the learning problem belongs to the class P.
Clearly, in this toy example, computational aspects are difficult to appreciate; however, these aspects are crucial in the real case. The more the parameters are the more expensive it is to explore the parameter space in order to find the “best” solution. Of course, a line can be represented by only two parameters while a curve requires (in general) more parameters.
This error is referred to as false negative.
References
Bishop CM (1995) Neural networks for pattern recognition. Oxford University Press, Oxford
Blumer A, Ehrenfeucht A, Haussler D, Warmuth MK (1989) Learnability and the Vapnik-Chervonenkis dimension. J ACM 36(4):929–965
Brock O, Khatib O (2002) Elastic strips: a framework for motion generation in human environments. Int J Robot Res 21(12):1031–1052
Capurro R, Nagenborg M, Weber J, Pingel C (2006) Methodological issues in the ethics of human-robot interaction. In: Tamburrini G, Datteri E (eds) Ethics of human interaction with robotic, bionic, and AI systems, workshop book of abstracts. Istituto Italiano per gli Studi Filosofici, Napoli, p 9
Cordeschi R, Tamburrini G (2005) Intelligent machinery and warfare: historical debates and epistemologically motivated concerns. In: Magnani L, Dossena R (eds) Computing, philosophy, and cognition. King’s College Publications, London, pp 1–23
Cucker F, Smale S (2001) On the mathematical foundations of learning. Bull Am Math Soc 39(1):1–49
Datteri E, Hosni H, Tamburrini G (2006) An inductionless and default-based analysis of machine learning procedures. In: Magnani L (eds) Model based reasoning in science and engineering. College Publications, London, pp 379–399
Fawcett T (2004) ROC graphs: notes and practical considerations for researchers. Tech report HPL-2003–4, HP Laboratories
Haussler D (1990) Probably approximately correct learning. In: AAAI-90 proceedings of the eight national conference on artificial intelligence, 1101–1108
Kohout R (2000) Challenges in real-time obstacle avoidance. In: AAAI spring symposium on real-time autonomous systems. Palo Alto
Matthias A (2004) The responsibility gap: ascribing responsibility for the actions of learning automata. Ethics Inf Technol 6:175–183
Minguez J, Montano L (2004) Nearness diagram navigation (ND): collision avoidance in troublesome scenarios. IEEE Trans Rob Autom 20(1):45–59
Mitchell TM (1997) Machine learning. McGraw Hill New York
Müller KR, Mika S, Ratsch G, Tsuda K, Scholkopf B (2001) An introduction to kernel-based learning algorithms. IEEE Trans Neural Netw 12(2):181–201
Pineau J, Gordon G (2005) “POMDP planning for robust robot control”. International Symposium on Robotics Research (ISRR), San Francisco
Pitt L, Valiant L (1988) Computational limitations on learning from examples. J ACM 35:965–984
Tamburrini G (2006) AI and Popper’s solution to the problem of induction. In: Jarvie I, Milford K, Miller D (eds) Karl popper: a centennial assessment, metaphysics and epistemology. Ashgate, London, 2:265–282
Valiant L (1984) A theory of the learnable. Commun ACM 27:1134–1142
Vapnik VN (1999a) An overview of statistical learning theory. IEEE Trans Neural Netw 10(5)
Vapnik VN (1999b) The nature of statistical learning theory. Springer, Heidelberg
Vapnik VN, Chervonenkis AY (1971) On the Uniform Convergence of relative frequencies of events to their probabilities. Theor Probab Appl 16:264–280
Vidyasagar M (1996) A theory of learning and generalization. Springer, Heidelberg
Zweig MH, Campbell G (1993) Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 39(4):561–577
Acknowledgements
The financial support of the VI Framework Programme of the European Union (Science and Society, contract number 017759) is gratefully acknowledged.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Santoro, M., Marino, D. & Tamburrini, G. Learning robots interacting with humans: from epistemic risk to responsibility. AI & Soc 22, 301–314 (2008). https://doi.org/10.1007/s00146-007-0155-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00146-007-0155-9