Logic-Guided Eligibility Traces for Delayed and Sparse Reward Reinforcement Learning

Author(s): Ravi Dayani

Publication #: 2602016

Date of Publication: 09.03.2026

Country: United States

Pages: 1-5

Published In: Volume 12 Issue 2 March-2026

Abstract

Delayed and sparse rewards present a fundamental challenge in reinforcement learning, degrading performance due to impaired temporal credit assignment. Traditional eligibility traces and multi-step returns propagate learning signals backward in time but implicitly assume timely reward observation, limiting effectiveness under delayed feedback. We propose Logic-Guided Eligibility Traces (LGET), a neuro-symbolic framework that integrates symbolic logical inference into the eligibility trace mechanism. A lightweight Prolog-based module infers relational relevance between past transitions and delayed rewards, and this relevance modulates trace updates, guiding learning toward causally pertinent events. We derive the LGET algorithm and establish convergence guarantees under bounded reward delays in tabular and linear approximation settings. Experiments on delayed and sparse reward benchmarks demonstrate faster convergence, improved stability, and higher sample efficiency compared to conventional actor–critic methods. These results highlight the potential of combining symbolic reasoning with neural learning dynamics to address challenging reinforcement learning scenarios.

Keywords: Reinforcement learning, Eligibility traces, Delayed rewards, Sparse rewards, Neuro-symbolic learning

Download/View Paper's PDF

Download/View Count: 2

Share this Article