Bounded autonomy: autonomy and dependence

“The agents have bounded autonomy”. What could this mean? Let us try to contribute to an answer with the help of (Conte and Castelfranchi 1995: Chps. 2 and 3).

To be autonomous an agent must be capable of generate new goals as means for achieving existing goals of its own. But, except for heavenly beings, autonomous agents are not self-sufficient. The autonomy is limited by dependence. An agent depends on a resource when he needs it to perform some action to achieve one of his goals. Beyond resource dependence, there is social dependence: an agent x depends on another agent y when, to achieve one of his goals, x needs an action of y. Agents can even treat other agents as resources. There is mutual dependence between two agents when they depend on each other to achieve one and the same goal. Dependences imply interests. A world state that favours the achievement of an agent’s goals is an interest of that agent.

The relations of dependence and interest hold whether an agent is aware of them or not. Objective relations between two or more agents or between agents and the external world are those relations that could be described by a non-participant observer even if they are not in the participants minds. So, there is an objective base of social interaction. There is social interference between two agents when the achievement of one’s goals has some (positive or negative) effects on the other achieving his goals – be those effects intended or unintended by any agent.
Limited autonomy of social agents comes also from influencing relations between them. By acquiring beliefs about their interests agents can acquire goals. An agent can have true beliefs about his interests, when they overlap with objective interests. True beliefs about interests can help setting goals and planning action. But an agent can also have false beliefs about interests, as well as ignoring some of his objective interests. Furthermore, there can be conflicting interests of the same agent (viz immediate vs. long-term interests).

Now, an agent can adopt another agent’s goals. If y has a goal g and x wants y to achieve g as long as x believes that y wants to achieve g, we can say that x adopted the y’s goal. The goal adoption can be a result of influencing: y can work to have x adopting some of y’s goals. By influencing, new goals can replace older ones. An agent x can influence another agent y to adopt a goal g according to x’s needs, even if that goal g is not an interest of y.

So, the bounded autonomy of the agents comes from the relations of dependence and influencing holding among them, and between them and the real world.

REFERENCE
(Conte and Castelfranchi 1995) CONTE, Rosaria, and CASTELFRANCHI, Cristiano, Cognitive and Social Action, London, The University College London Press, 1995

Read more...

Autonomy

Pfeifer and Bongard (2007), dealing with design principles for collective systems, suggest that, according to the “level of abstraction principle”, collective intelligence refers not only to groups of individuals, as in human societies, but equally “to any kind of assembly of similar agents”, including groups of modules in modular robotic systems or organs that make up entire organisms (Pfeifer and Bongard 2007:241-243). Now, the “level of abstraction principle” raises the following question: to put individuals (for example) in human societies on the same foot with organs or modules purports to ignore different degrees of autonomy enjoyed by a human lung and a human individual. Pim Haselager helps to elaborate on that question.

According to Haselager, the following definition sums up various interpretations of autonomous agents circulating within AI: “Autonomous agents operate under all reasonable conditions without recourse to an outside designer, operator or controller while handling unpredictable events in an environment or niche” (Haselager 2005:518). This could be a working definition within robotics, relating more autonomy to less intervention of human beings while the robot is operating, and ruling out completely predetermined environments.
However, from some philosophical perspectives this conception of autonomy would be unsatisfactory, because it lacks an appropriate emphasis on the reasons for acting. A truly autonomous agent must be capable of acting according to her own goals and choices, while robots don’t choose their goals. Programmers and designers are the sole providers of goals to the robots. Notwithstanding, roboticists can safely ignore this “free-will concept of autonomy”. Mechanistic inclined philosophers do the same. For them, free-will is just an illusion and even adult human beings have no real choices.

Haselager offers a third concept of autonomy that could narrow the gap between autonomy-in-robotics and autonomy-in-philosophy. This concept of autonomy focus on homeostasis and the intrinsic ownership of goals.
A system can have his own goals, even if it cannot freely choose them, if they matter to his success or failure. A robot owns his goals “when they arise out of the ongoing attempt, sustained by both the body and the control system, to maintain homeostasis” (Haselager 2005:523). For example, a robot regulating his level of energy is in some way aiming for a goal of his own. This is still true despite the fact the robot is not free to ignore that specific goal. Evolutionary robotics, allowing the human programmer to withdraw from the design of that behaviour, still increases autonomy. That approach could be further improved with co-evolution of body and control systems, as much as adding autopoiesis to homeostasis. Notwithstanding, our understanding of autonomy, both in technical and in philosophical terms, could benefit from those ways to experiment how goals become grounded in artificial creatures.

Whether full autonomy is attainable is a remaining question.


REFERENCES

(Pfeifer and Bongard 2007) PFEIFER, R., and BONGARD, J., How the Body Shapes the Way We Think, Cambridge: Massachusetts, The MIT Press, 2007

(Haselager 2005) HASELAGER, Willem F.G., “Robotics, philosophy and the problems of autonomy”, in Pragmatics & Cognition, 13(3), 515-532 (click to open the pdf file)

Haselager page

Read more...