A learning system based on lazy metareasoning

摘要: Metareasoning has been widely studied in the literature, with a wide variety of algorithms and partially overlapping methodological approaches. However, these methods are typically either not targeted toward practical machine learning systems or alternatively focused on achieving best possible performance for particular domain, extensive human tuning research, vast computing resources. In this paper, our goal is to create that perform sustained autonomous learning, automatically determined domain-specific optimizations any given without requiring assistance. We present Alma, metareasoning architecture creates selects reasoning based empirically observed performance. This achieved by using lazy at metalevel, training combining run-time. experiments across diverse data sets, we demonstrate ability Alma successfully reason about learner different domains achieve better overall result than individual methods, even limited time available.