Risk-averse Total-reward MDPs with ERM and EVaR

Authors: Xihong Su, Marek Petrik, Julien Grand-Clément

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate the effect of risk-aversion on the structure of the optimal policy, we use the gambler s ruin problem (Hau, Petrik, and Ghavamzadeh 2023; B auerle and Ott 2011). In this problem, a gambler starts with a given amount of capital and seeks to increase it up to a cap K. ... The algorithm was implemented in Julia 1.10, and is available at https://github.com/suxh2019/ERMLP. Please see Su, Grand Cl ement, and Petrik (2024, appendix F) for more details. Figure 3 shows optimal policies for four different EVa R risk levels α computed by Algorithm 1. ... To understand the impact of risk-aversion on the distribution of returns, we simulate the resulting policies over 7,000 episodes and show the distribution of capitals in Figure 4.
Researcher Affiliation Academia 1University of New Hampshire, 33 Academic Way, Durham, NH, 03824 USA 2 HEC Paris, 1 Rue de la Lib eration, Jouy-en-Josas, 78350 France EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Simple EVa R algorithm
Open Source Code Yes The algorithm was implemented in Julia 1.10, and is available at https://github.com/suxh2019/ERMLP.
Open Datasets No The paper uses the 'gambler's ruin problem' as a simulation environment with specified parameters (q=0.68, K=7) and simulates policies over 7,000 episodes. This is a self-generated simulated dataset, not a publicly available external dataset with access information.
Dataset Splits No The paper describes a simulation study for the gambler's ruin problem where policies are simulated over 7,000 episodes. It does not mention any explicit training, validation, or test splits for a dataset, as the data is generated through simulation rather than being a pre-existing dataset that needs splitting.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions the implementation language.
Software Dependencies Yes The algorithm was implemented in Julia 1.10
Experiment Setup Yes In the formulation, we use q = 0.68, and a cap is K = 7. ... Figure 3 shows optimal policies for four different EVa R risk levels α computed by Algorithm 1. ... To understand the impact of risk-aversion on the distribution of returns, we simulate the resulting policies over 7,000 episodes