Noise Tolerance in Reinforcement Learning Algorithms

Abstract

Referrals are used in multi-agent systems, network agents and peer-to-peer systems for the purpose of global or local information spreading to facilitate trust relationships and reciprocal interactions. Based on referral local interactions can be altered with a purpose to maximise the utility function of each of the participants, which in many cases requires mutual co-operation of participants. The referral system is often based on the global or statistical behaviour of the overall society. In this article, we provide a simple taxonomy of referral systems and on that basis we discuss three distinct ways information can be collected and aggregated. We analyse the effects of global vs. local information spreading, in terms of individual agent and global (societal) performance of a population based on the maximisation of a utility function of each of the agents. Our studies show that under certain conditions such as large number of non uniformly acting autonomous agents the spread of global information is undesirable. Collecting and providing local information only yields better overall results. In some experimental setups however, it might be necessary for global information to be available otherwise global stable optimal behaviour cannot be achieved. We analyse both of these extreme cases based on simple game-theoretic setup. We analyse and relate our results in the context of e-mail relying and spam filtering^1.