Universal semantic communication

Abstract

Is meaningful communication possible between two intelligent parties who share no common language or background? We propose that this problem can be rigorously addressed by explicitly focusing on the goals of the communication. We propose a theoretical framework in which we can address when and to what extent such semantic communication is possible. Our starting point is a mathematical definition of a generic goal for communication, that is pursued by agents of bounded computational complexity. We then model a "lack of common language or background" by considering a class of potential partners for communication; in general, this formalism is rich enough to handle varying degrees of common language and backgrounds, but the complete lack of knowledge is modeled by simply considering the class of all partners with which some agent of similar power could achieve our goal. In this formalism, we will find that for many goals (but not all), communication without any common language or background is possible. We call the strategies for achieving goals without relying on such background universal protocols. The main intermediate notions introduced by our theory are formal notions of feedback that we call sensing. We show that sensing captures the essence of whether or not reliable universal protocols can be constructed in many natural settings of interest: we find that across settings, sensing is almost always sufficient, usually necessary, and generally a useful design principle for the construction of universal protocols. We support this last point by developing a number of examples of protocols for specific goals. Notably, we show that universal delegation of computation from a space-efficient client to a general-purpose server is possible, and we show how a variant of TCP can allow end-users on a packet network to automatically adapt to small changes in the packet format (e.g., changes in IP). The latter example above alludes to our main motivation for considering such problems, which is to develop techniques for modeling and constructing computer systems that do not require that their components strictly adhere to protocols: said differently, we hope to be able to design components that function properly with a sufficiently wide range of other components to permit a rich space of "backwards-compatible" designs for those components. We expect that in the long run, this paradigm will lead to simpler systems because "backwards compatibility" is no longer such a severe constraint, and we expect it to lead to more robust systems, partially because the components should be simpler, and partially because such components are inherently robust to deviations from any fixed protocol. Unfortunately, we find that the techniques for communication under the complete absence of any common background suffer from overhead that is too severe for such practical purposes, so we consider two natural approaches for introducing some assumed common background between components while retaining some nontrivial amount of flexibility. The first approach supposes that the designer of a component has some "belief" about what protocols would be "natural" to use to interact with other components; we show that, given sensing and some sufficient "agreement" between the beliefs of the designers of two components, the components can be made universal with some relatively modest overhead. The second approach supposes that the protocols are taken from some restricted class of functions, and we will see that for certain classes of functions and simple goals, efficient universal protocols can again be constructed from sensing. Actually, we show more: the special case of our model described in the second approach above corresponds precisely to the well-known model of mistake-bounded on-line learning first studied by Barzdirs and Frievalds, and later considered in more depth by Littlestone. This connection provides a reasonably complete picture of the conditions under which we can apply the second approach. Furthermore, it also seems that the first approach is closely related to the problem of designing good user interfaces in Human-Computer Interaction. We conclude by briefly sketching the connection, and suggest that further development of this connection may be a potentially fruitful direction for future work.