122
I. T. Freire et al.
taxonomy of moral values, each of which is expressed as a solution to a distinct cooperation game. In a recent study this taxonomy has been empirically
validated through finding instances of these seven moral principles across 60
cultures [13].
5 Machine-Morality-as-Cooperation Through Distributed
Adaptive Control
Morality-as-Cooperation seems to be a promising framework for future human
studies, but is there any similar synthetic approach addressing moral problems in
robotics? From our literature review it seems that the widely accepted angle so
far has been the one mirroring the harm aversion perspective in human morality
studies, namely harm avoidance: a machine that is moral is defined as the one
that does no harm to humans. While acknowledging its utility, we could argue
that harm avoidance is a condition necessary but not sufficient to call a machine
moral, and our cross-disciplinary review has justified the proposal of a new perspective that could be fruitful in the development of Human-Robot Interaction
scenarios: Machine-Morality-as-Cooperation.
Machine-Morality-as-Cooperation (MMAC) is a synthetic and functional
approach to modelling machine morality as it incorporates the previous principles adopted in the field -harm avoidance- while also adding a novel dimension
grounded in human morality studies: promote cooperation. At the core of this
approach lies the implementation of a reactive control system that, inspired by
the Pavlovian appetitive and aversive reactive drives, will apply specific rules
to ensure both harm avoidance and prosocial cooperative behaviors. Moreover,
in order to support both deontological and consequentalist computations, this
artificial system should incorporate both model-free and model-based forms of
learning. Computational single-architecture integration of the three cognitive
networks aimed at implementing Machine-Morality-as-Cooperation uses the Distributed Adaptive Control (DAC) design principles. Several realisations of the
DAC architecture have been previously successfully applied and validated on different robotic platforms [26,29,40], and therefore DAC can be a good candidate
to instantiate Machine-Morality-as-Cooperation requirements in an autonomous
agent.
6 Machine-Morality-as-Cooperation: Prescriptive
and Functional Implementations
As with the study of morality, we can discuss the application of MachineMorality-as-Cooperation from a prescriptive, high-level standpoint, or choose
a more grounded and functional way. Both cases deserve attention due to their
theoretical and practical implications. In order to realise an ideal implementation
of MMAC’s cognitive architecture in an artificial autonomous agent, the agent
will first have to pass through three specific developmental stages -according to
the requirements of each layer- in order to reach its full capacity.
I. T. Freire et al.
taxonomy of moral values, each of which is expressed as a solution to a distinct cooperation game. In a recent study this taxonomy has been empirically
validated through finding instances of these seven moral principles across 60
cultures [13].
5 Machine-Morality-as-Cooperation Through Distributed
Adaptive Control
Morality-as-Cooperation seems to be a promising framework for future human
studies, but is there any similar synthetic approach addressing moral problems in
robotics? From our literature review it seems that the widely accepted angle so
far has been the one mirroring the harm aversion perspective in human morality
studies, namely harm avoidance: a machine that is moral is defined as the one
that does no harm to humans. While acknowledging its utility, we could argue
that harm avoidance is a condition necessary but not sufficient to call a machine
moral, and our cross-disciplinary review has justified the proposal of a new perspective that could be fruitful in the development of Human-Robot Interaction
scenarios: Machine-Morality-as-Cooperation.
Machine-Morality-as-Cooperation (MMAC) is a synthetic and functional
approach to modelling machine morality as it incorporates the previous principles adopted in the field -harm avoidance- while also adding a novel dimension
grounded in human morality studies: promote cooperation. At the core of this
approach lies the implementation of a reactive control system that, inspired by
the Pavlovian appetitive and aversive reactive drives, will apply specific rules
to ensure both harm avoidance and prosocial cooperative behaviors. Moreover,
in order to support both deontological and consequentalist computations, this
artificial system should incorporate both model-free and model-based forms of
learning. Computational single-architecture integration of the three cognitive
networks aimed at implementing Machine-Morality-as-Cooperation uses the Distributed Adaptive Control (DAC) design principles. Several realisations of the
DAC architecture have been previously successfully applied and validated on different robotic platforms [26,29,40], and therefore DAC can be a good candidate
to instantiate Machine-Morality-as-Cooperation requirements in an autonomous
agent.
6 Machine-Morality-as-Cooperation: Prescriptive
and Functional Implementations
As with the study of morality, we can discuss the application of MachineMorality-as-Cooperation from a prescriptive, high-level standpoint, or choose
a more grounded and functional way. Both cases deserve attention due to their
theoretical and practical implications. In order to realise an ideal implementation
of MMAC’s cognitive architecture in an artificial autonomous agent, the agent
will first have to pass through three specific developmental stages -according to
the requirements of each layer- in order to reach its full capacity.