Speaker
Description
State-space models (SSMs) have emerged as a compelling
alternative to Transformer architectures, delivering comparable
performance at significantly lower computational cost. Al-
though deterministic SSMs such as Mamba have achieved
state-of-the-art results in areas like sequence modelling and
image segmentation, their deterministic nature limits their
suitability for probabilistic reinforcement learning (RL) en-
vironments, where uncertainty is intrinsic.
This paper introduces an architecture that integrates the
Mamba SSM with the Kalman filter, enabling it to learn and
adapt to uncertain environment dynamics. We validate the ef-
fectiveness of this approach on a modified Pendulum task from
the Gymnasium RL library, demonstrating its potential for
learning and representing complex dynamics in probabilistic
settings.
