Learning structure for
value and exploration.

The research question

How can an agent learn useful structural representations of an environment from experience? My PhD research investigates Laplacian representations of environment dynamics and their use in value function approximation and exploration in deep reinforcement learning.

Value learning with Laplacian representations

“Revisiting Laplacian Representations for Value Function Approximation in Deep RL” examines these representations as a basis for approximating value functions. It was presented at the Inductive Biases in Reinforcement Learning Workshop at RLC 2025.

Exploration with hierarchical eigen options

“Addressing Sparse-Rewards in RL with Scalable Hierarchical Novel Eigen Options” studies scalable hierarchical options for exploration in sparse-reward environments. It is accepted at NeurIPS 2026 and was also presented as a spotlight at the Reinforcement Learning Beyond Rewards Workshop at RLC 2026.

A shared foundation

These two strands use spectral representations to connect learned environment structure with decision-making: one studies value function approximation; the other studies temporally extended exploration.

Related publications & artifacts

2026

Addressing Sparse-Rewards in RL with Scalable Hierarchical Novel Eigen Options

P Vijayan, E C Gauthier, M Reymond, S Chandar, D Precup, and I P Schwarz

Fortieth Annual Conference on Neural Information Processing Systems, NeurIPS'26

Reinforcement Learning Beyond Rewards Workshop, RLC 2026

2025

Revisiting Laplacian Representations for Value Function Approximation in Deep RL

P Vijayan, P Nouri, R Rishav, S Chandar, Y Chandak, M Reymond, S E Kahou, and D Precup

Inductive Biases in Reinforcement Learning Workshop, RLC 2025

Paper ↗