Semantic Communication for Simple Goals Is Equivalent to On-line Learning
Previous works [11,6] introduced a model of semantic communication between a “user” and a “server,” in which the user attempts to achieve a given goal for communication. They show that whenever the user can sense progress, there exist universal user strategies that can achieve the goal whenever it is possible for any other user to reliably do so. A drawback of the actual constructions is that the users are inefficient: they enumerate protocols until they discover one that is successful, leading to the potential for exponential overhead in the length of the desired protocol. Goldreich et al.  conjectured that this overhead could be reduced to a polynomial dependence if we restricted our attention to classes of sufficiently simple user strategies and goals. In this work, we are able to obtain such universal strategies for some reasonably general special cases by establishing an equivalence between these special cases and the usual model of mistake-bounded on-line learning [3,15]. This equivalence also allows us to see the limits of constructing universal users based on sensing and motivates the study of sensing with richer kinds of feedback. Along the way, we also establish a new lower bound for the “beliefs model” , which demonstrates that constructions of efficient users in that framework rely on the existence of a common “belief” under which all of the servers in a class are designed to be efficient.
Keywordssemantic communication on-line learning feedback models
Unable to display preview. Download preview PDF.
- 3.Bārzdiņš, J., Freivalds, R.: On the prediction of general recursive functions. Soviet Math. Dokl. 13, 1224–1228 (1972)Google Scholar
- 6.Goldreich, O., Juba, B., Sudan, M.: A theory of goal-oriented communication. Tech. Rep. TR09-075, ECCC (2009)Google Scholar
- 7.Grötschel, M., Lovász, L., Schrijver, A.: Geometric methods in combinatorial optimization. In: Pulleybank, W.R. (ed.) Proc. Silver Jubilee Conf. on Combinatorics. Progress in Combinatorial Optimization, pp. 167–183. Academic Press, New York (1984)Google Scholar
- 9.Jayanti, P., Toueg, S.: Every problem has a weakest failure detector. In: 27th PODC (2008)Google Scholar
- 10.Juba, B.: Universal Semantic Communication. Ph.D. thesis, MIT (2010)Google Scholar
- 11.Juba, B., Sudan, M.: Universal semantic communication I. In: 40th STOC (2008)Google Scholar
- 12.Juba, B., Sudan, M.: Efficient semantic communication via compatible beliefs. In: 2nd Innovations in Computer Science (2011)Google Scholar
- 14.Kharitonov, M.: Cryptographic hardness of distribution-specific learning. In: 25th STOC, pp. 372–381 (1993)Google Scholar
- 15.Littlestone, N.: Learning quickly when irrelevant attributes abound: A new linear-threshold algorithm. Mach. Learn. 2(4), 285–318 (1988)Google Scholar
- 16.Maass, W., Turán, G.: How fast can a threshold gate learn? In: Hanson, S.J., Drastal, G.A., Rivest, R.L. (eds.) Computational Learning Theory and Natural Learning Systems: Constraints and Prospects, vol. 1, pp. 381–414. MIT Press, Cambridge (1994)Google Scholar
- 17.Patt-Shamir, B.: Personal communication (2010)Google Scholar