Discovering a set of policies for the worst case reward

Authors: Tom Zahavy, Andre Barreto, Daniel J Mankowitz, Shaobo Hou, Brendan O'Donoghue, Iurii Kemaev, Satinder Singh

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate our algorithm on a grid world and also on a set of domains from the Deep Mind control suite. We confirm our theoretical results regarding the monotonically improving performance of our algorithm.
Researcher Affiliation Industry Tom Zahavy , Andre Barreto, Daniel J Mankowitz, Shaobo Hou, Brendan O Donoghue, Iurii Kemaev and Satinder Singh Deep Mind
Pseudocode Yes Algorithm 1 SMP worst case policy iteration
Open Source Code No The paper does not include an unambiguous statement about releasing source code for the described methodology or a direct link to a source-code repository.
Open Datasets Yes We empirically evaluate our algorithm on a grid world and also on a set of domains from the Deep Mind control suite. Next, we conducted a set of experiments in the DM Control Suite (Tassa et al., 2018).
Dataset Splits No The paper describes training policies and evaluating them, but it does not specify explicit dataset splits (e.g., percentages or counts) for training, validation, and testing.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions specific software components like "STACX (Zahavy et al., 2020d)" and "CVXPY (Diamond & Boyd, 2016)" but does not provide version numbers for them.
Experiment Setup Yes At each iteration (x-axis) of Algorithm 1 we train a policy for 5 105 steps to maximize w SMP Πt . We then compute the SFs of that policy using additional 5 105 steps and evaluate it w.r.t w SMP Πt . At each iteration of Algorithm 1 we train a policy for 2 106 steps using an actor-critic (and specifically STACX (Zahavy et al., 2020d)) to maximize w SMP Πt , add it to the set, and compute a new w SMP Πt+1. We focused on the setup where the agent is learning from feature observations corresponding to the positions and velocities of the body in the task (pixels were only used for visualization).