یک معماری شناختی قوی برای یادگیری از شگفتی ها

Learning from surprises is a cornerstone for building bio-inspired cognitive architectures that can autonomously learn from interactions with their environments. However, distinguishing true surprises – from which useful information can be extracted to improve an agent’s world model – from environmental noise is a fundamental challenge. This paper proposes a new and robust approach for actively learning a predictive model of discrete, stochastic, partially-observable environments based on a concept called the Stochastic Distinguishing Experiment (SDE). SDEs are conditional probability distributions over the next observation given a variable-length sequence of ordered actions and expected observations up to the present that partition the space of possible agent histories, thus forming an approximate predictive representation of state. We derive this SDE-based learning algorithm and present theoretical proofs of its convergence and computational complexity. Theoretical and experimental results in small environments with important theoretical properties demonstrate the algorithm’s ability to build an accurate predictive model from one continuous interaction with its environment without requiring any prior knowledge of the underlying state space, the number of SDEs to use, or even a bound on SDE length.