Towards Efficient Risk-Sensitive Policy Gradient: An Iteration Complexity Analysis

Reinforcement Learning (RL) has shown exceptional performance across various applications, enabling autonomous agents to learn optimal policies through interaction with their environments. However, traditional RL frameworks often face challenges in terms of iteration complexity and robustness. Risk-sensitive RL, which balances expected return and risk, has been explored for its potential to yield probabilistically robust policies, yet its iteration complexity analysis remains underexplored. In this study, we conduct a thorough iteration complexity analysis for the risk-sensitive policy gradient method, focusing on the REINFORCE algorithm and employing the exponential utility function. We obtain an iteration complexity of $\mathcal{O}(\epsilon^{-2})$ to reach an $\epsilon$-approximate first-order stationary point (FOSP). We investigate whether risk-sensitive algorithms can achieve better iteration complexity compared to their risk-neutral counterparts. Our theoretical analysis demonstrates that risk-sensitive REINFORCE can have a reduced number of iterations required for convergence. This leads to improved iteration complexity, as employing the exponential utility does not entail additional computation per iteration. We characterize the conditions under which risk-sensitive algorithms can achieve better iteration complexity. Our simulation results also validate that risk-averse cases can converge and stabilize more quickly after approximately half of the episodes compared to their risk-neutral counterparts.

Non-ergodicity in reinforcement learning: robustness via ergodicity transformations

Envisioned application areas for reinforcement learning (RL) include autonomous driving, precision agriculture, and finance, which all require RL agents to make decisions in the real world. A significant challenge hindering the adoption of RL methods in these domains is the non-robustness of conventional algorithms. In this paper, we argue that a fundamental issue contributing to this lack of robustness lies in the focus on the expected value of the return as the sole "correct" optimization objective. The expected value is the average over the statistical ensemble of infinitely many trajectories. For non-ergodic returns, this average differs from the average over a single but infinitely long trajectory. Consequently, optimizing the expected value can lead to policies that yield exceptionally high returns with probability zero but almost surely result in catastrophic outcomes. This problem can be circumvented by transforming the time series of collected returns into one with ergodic increments. This transformation enables learning robust policies by optimizing the long-term return for individual agents rather than the average across infinitely many trajectories. We propose an algorithm for learning ergodicity transformations from data and demonstrate its effectiveness in an instructive, non-ergodic environment and on standard RL benchmarks.

Robust Counterfactual Explanations for Neural Networks With Probabilistic Guarantees

There is an emerging interest in generating robust counterfactual explanations that would remain valid if the model is updated or changed even slightly. Towards finding robust counterfactuals, existing literature often assumes that the original model $m$ and the new model $M$ are bounded in the parameter space, i.e., $\|\text{Params}(M){-}\text{Params}(m)\|{<}\Delta$. However, models can often change significantly in the parameter space with little to no change in their predictions or accuracy on the given dataset. In this work, we introduce a mathematical abstraction termed \emph{naturally-occurring} model change, which allows for arbitrary changes in the parameter space such that the change in predictions on points that lie on the data manifold is limited. Next, we propose a measure -- that we call \emph{Stability} -- to quantify the robustness of counterfactuals to potential model changes for differentiable models, e.g., neural networks. Our main contribution is to show that counterfactuals with sufficiently high value of \emph{Stability} as defined by our measure will remain valid after potential ``naturally-occurring'' model changes with high probability (leveraging concentration bounds for Lipschitz function of independent Gaussians). Since our quantification depends on the local Lipschitz constant around a data point which is not always available, we also examine practical relaxations of our proposed measure and demonstrate experimentally how they can be incorporated to find robust counterfactuals for neural networks that are close, realistic, and remain valid after potential model changes.

Risk-Sensitive Reinforcement Learning with Exponential Criteria

While risk-neutral reinforcement learning has shown experimental success in a number of applications, it is well-known to be non-robust with respect to noise and perturbations in the parameters of the system. For this reason, risk-sensitive reinforcement learning algorithms have been studied to introduce robustness and sample efficiency, and lead to better real-life performance. In this work, we introduce new model-free risk-sensitive reinforcement learning algorithms as variations of widely-used Policy Gradient algorithms with similar implementation properties. In particular, we study the effect of exponential criteria on the risk-sensitivity of the policy of a reinforcement learning agent, and develop variants of the Monte Carlo Policy Gradient algorithm and the online (temporal-difference) Actor-Critic algorithm. Analytical results showcase that the use of exponential criteria generalize commonly used ad-hoc regularization approaches. The implementation, performance, and robustness properties of the proposed methods are evaluated in simulated experiments.