Entropic Risk Optimization in Discounted MDPs: Sample Complexity Bounds with a Generative Model
This work addresses risk-sensitive reinforcement learning for agents with specific risk preferences, offering theoretical insights but is incremental in extending existing bounds to risk-aware settings.
The paper tackles the problem of learning optimal policies in discounted Markov decision processes with entropic risk preferences, providing sample complexity bounds for a model-based algorithm and showing that exponential dependence on risk sensitivity and horizon is unavoidable.
In this paper, we analyze the sample complexities of learning the optimal state-action value function $Q^*$ and an optimal policy $π^*$ in a finite discounted Markov decision process (MDP) where the agent has recursive entropic risk-preferences with risk-parameter $β\neq 0$ and where a generative model of the MDP is available. We provide and analyze a simple model based approach which we call model-based risk-sensitive $Q$-value-iteration (MB-RS-QVI) which leads to $(\varepsilon,δ)$-PAC-bounds on $\|Q^*-Q^k\|$, and $\|V^*-V^{π_k}\|$ where $Q_k$ is the output of MB-RS-QVI after k iterations and $π_k$ is the greedy policy with respect to $Q_k$. Both PAC-bounds have exponential dependence on the effective horizon $\frac{1}{1-γ}$ and the strength of this dependence grows with the learners risk-sensitivity $|β|$. We also provide two lower bounds which shows that exponential dependence on $|β|\frac{1}{1-γ}$ is unavoidable in both cases. The lower bounds reveal that the PAC-bounds are tight in the parameters $S,A,δ,\varepsilon$ and that unlike in the classical setting it is not possible to have polynomial dependence in all model parameters.