I'm pleased to announce a new paper from MIRI about The Value Learning Problem.
A superintelligent machine would not automatically act as intended: it will act as programmed, but the fit between human intentions and formal specification could be poor. We discuss methods by which a system could be constructed to learn what to value. We highlight open problems specific to inductive value learning (from labeled training data), and raise a number of questions about the construction of systems which model the preferences of their operators and act accordingly.
This is the sixth of six papers supporting the MIRI technical agenda. It motivates the need for value learning, a bit, and gives some early thoughts on how the problem could be approached (while pointing to some early open problems in the field).
I'm pretty excited to have the technical agenda and all its supporting papers published. Next week I'll be posting an annotated bibliography that gives more reading for each subject. The introduction to the value learning paper has been reproduced below.
Consider a superintelligent system, in the sense of Bostrom (2014), tasked with curing cancer by discovering some process which eliminates cancerous cells from a human body without causing harm to the human (no easy task to specify in its own right). The resulting behavior may be quite unsatisfactory. Among the behaviors not ruled out by this goal specification are stealing resources, proliferating robotic laboratories at the expense of the biosphere, and kidnapping human test subjects.
The intended goal, hopefully, was to cure cancer without doing any of those things, but computer systems do not automatically act as intended. Even a system smart enough to figure out what was intended is not compelled to act accordingly: human beings, upon learning that natural selection ``intended" sex to be pleasurable only for purposes of reproduction, do not thereby conclude that contraceptives are abhorrent. While one should not anthropomorphize natural selection, humans are capable of understanding the process which created them while being unmotivated to alter their preferences accordingly. For similar reasons, when constructing an artificially intelligent system, it is not sufficient to construct a system intelligent enough to understand human intentions; the system must also be purposefully constructed to pursue them (Bostrom 2014, chap. 8).
How can this be done? Human goals are complex, culturally laden, and context-dependent. Furthermore, the notion of ``intention" itself may not lend itself to clean formal specification. By what methods could an intelligent machine be constructed to reliably learn what to value and to act as its operators intended?
A superintelligent machine would be useful for its ability to find plans that its programmers never imagined, to identify shortcuts that they never noticed or considered. That capability is a double-edged sword: a machine that is extraordinarily effective at achieving its goals might have unexpected negative side effects, as in the case of robotic laboratories damaging the biosphere. There is no simple fix: a superintelligent system would need to learn detailed information about what is and isn't considered valuable, and be motivated by this knowledge, in order to safely solve even simple tasks.
This value learning problem is the focus of this paper. Section 2 discusses an apparent gap between most intuitively desirable human goals and attempted simple formal specifications. Section 3 explores the idea of frameworks through which a system could be constructed to learn concrete goals via induction on labeled data, and details possible pitfalls and early open problems. Section 4 explores methods by which systems could be built to safely assist in this process.
Given a system which is attempting to act as intended, philosophical questions arise: How could a system learn to act as intended when the operators themselves have poor introspective access to their own goals and evaluation criteria? These philosophical questions are discussed briefly in Section 5.
A superintelligent system under the control of a small group of operators would present a moral hazard of extraordinary proportions. Is it possible to construct a system which would act in the interests of not only its operators, but of all humanity, and possibly all sapient life? This is a crucial question of philosophy and ethics, touched upon only briefly in Section 6, which also motivates a need for caution and then concludes.