Agent-Specific Effects: A Causal Effect Propagation Analysis in Multi-Agent MDPs
Authors: Stelios Triantafyllou, Aleksa Sukovic, Debmalya Mandal, Goran Radanovic
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we experimentally evaluate the utility of cf ASE through a simulation-based testbed, which includes a sepsis management environment. and We conduct extensive experiments on two test-beds, Graph and Sepsis, evaluating the robustness and practicality of our approach. |
| Researcher Affiliation | Academia | 1Max Planck Institute for Software Systems, Saarbr ucken, Germany 2Saarland University, Saarbr ucken, Germany 3University of Warwick, Department of Computer Science, UK. |
| Pseudocode | Yes | Algorithm 1 Estimates cf-ASEN ai,t(y|τ)M |
| Open Source Code | Yes | 1Code to reproduce our experiments is available at https://github.com/stelios30/agent-specific-effects.git. |
| Open Datasets | Yes | In this section, we empirically evaluate our approach using two environments from prior work, Graph (Triantafyllou et al., 2021) and Sepsis (Oberst & Sontag, 2019), modified to align with our objectives. |
| Dataset Splits | No | The paper describes generating and selecting trajectories for evaluation ('For each environment, we generate a set of trajectories in which agents fail to reach their goal, 500 for Graph and 100 (per value of µ) for Sepsis. For each trajectory τ, we compute the TCFE of all the potential alternative actions that agents could have taken. Among these actions ai,t, we retain those that exhibit a TCFE greater than or equal to a predefined threshold...') but does not explicitly specify train/validation/test dataset splits for the main experiments, nor mention cross-validation. |
| Hardware Specification | Yes | All experiments were run on a 64bit Debian-based machine having 4x12 CPU cores clocked at 3GHz with access to 1.5TB of DDR3 1600MHz RAM. |
| Software Dependencies | Yes | The software stack relied on Python 3.9.13, with installed standard scientific packages for numeric calculations and visualization: Num Py (1.24.3), Pandas (2.0.3), Joblib (1.3.1), Matplotlib (3.7.1) and Seaborn (0.12.2). For learning agent s policies in sepsis experiments, we relied on Py MDPToolbox (4.0-b3) library. |
| Experiment Setup | Yes | For each trajectory τ, we compute the TCFE of all the potential alternative actions that agents could have taken. Among these actions ai,t, we retain those that exhibit a TCFE greater than or equal to a predefined threshold θ (0.75 for Graph and 0.8 for Sepsis). ... Counterfactual effects in our experiments are computed for 100 counterfactual samples. |