{"id":"arxiv/2603.26464","url":"https://arxiv.org/abs/2603.26464","tags":["cs.LG","math.DS"],"score":0,"title":"Automatic feature identification in least-squares policy iteration using the Koopman operator framework","domain":"cs.LG","source":"arxiv","authors":["Christian Mugisho Zagabe","Sebastian Peitz"],"reviews":0,"subject":"cs.LG","summary":"In this paper, we present a Koopman autoencoder-based least-squares policy iteration (KAE-LSPI) algorithm in reinforcement learning (RL). The KAE-LSPI algorithm is based on reformulating the so-called least-squares fixed-point approximation method in terms of extended dynamic mode decomposition (EDMD), thereby enabling automatic feature learning via the Koopman autoencoder (KAE) framework. The approach is motivated by the lack of a systematic choice of features or kernels in linear RL techniques. We compare the KAE-LSPI algorithm with two previous works, the classical least-squares policy…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-03-27","reactions":{}}