The role of simulation in machine learning research

Abstract

This paper discusses the role of simulation in machine learning studies and presents a view of simulation-based machine learning. Based on the concept of the intelligent agent, it is shown how each of a variety of learning subsystems interacts with a simulated performance engine and how they may interact with each other. In particular, in the context of ongoing research into the coordination of various approaches to learning into an integrated facility called the Learning Testbed, the centrality of the simulation performance engine NETSIM to the development of the Learning Testbed is discussed. NETSIM is a fine-grained simulation of the call placement process in a circuit-switched telecommunications network which allows observation of the effectiveness of various traffic control strategies on network performance when time-varying traffic patterns are encountered. The users of the NETSIM program are three learning programs, which embody three different approaches to how a specialized domain, such as network traffic control, might be learned.