Explainability Via Causal Self-Talk
Authors: Nicholas A. Roy, Junkyung Kim, Neil Rabinowitz
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We implement this method in a simulated 3D environment, and show how it enables agents to generate faithful and semantically-meaningful explanations of their own behavior. |
| Researcher Affiliation | Industry | Nicholas A. Roy Deep Mind nroy@deepmind.com Junkyung Kim Deep Mind junkyung@deepmind.com Neil Rabinowitz Deep Mind ncr@deepmind.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Our agent training code is built on a core library that has yet to be open-sourced. If this library is published prior to Neur IPS camera ready deadline, we will endeavor to open-source the code for the design presented in this paper in the camera-ready version. |
| Open Datasets | No | We study variants of CST in a 3D virtual environment built in Unity [1, 68]. This features a fixed-layout indoor space comprising five rooms, each with a different wall color. We developed a simple task in this environment, called Dax Ducks (Figure 2). The paper describes a custom simulated environment and task without providing public access information for the data generated or the environment itself. |
| Dataset Splits | No | The paper describes training schedules and evaluation methodologies (e.g., 'evaluation episodes', 'simulated interventions in replay') but does not provide explicit data splits (e.g., percentages or sample counts) for training, validation, and test sets. |
| Hardware Specification | Yes | All training was performed on a cluster of custom-built machines, each comprising an Intel Xeon W-2295 CPU and 256GB of RAM, and using NVIDIA V100 or A100 GPUs (with 32GB or 40GB of memory). Each agent typically trained using 32 to 64 CPU cores and between 1 and 8 GPUs. |
| Software Dependencies | No | The paper mentions software components like 'Unity [1, 68]' and 'V-trace [22]' but does not provide specific version numbers for these or any other software dependencies needed to replicate the experiments. |
| Experiment Setup | Yes | We used the following schedules for interventions at training time. For CST-RL, interventions (mt m0) occurred randomly, with probability p = 0.03 that an intervention would occur at any given timestep t. For CST-MR, we simulated interventions in replay at every timestep. For CST-PD, we simulated interventions in replay: we divided every trajectory into a sequence of blocks of variable duration, with a p = 0.03 probability that a new block would start at any time t; we computed LP D only up to a horizon of the end of the block, and with a constant discounting function, γ( t) = 1. All agents were trained for 2B (2 × 10^9) environment steps. |