Inferring the Future by Imagining the Past
Authors: Kartik Chandra, Tony Chen, Tzu-Mao Li, Jonathan Ragan-Kelley, Josh Tenenbaum
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 2 we formalize this problem and present a new Monte Carlo algorithm for sampling approximate solutions. Then, in Section 3, we show that our algorithm is up to 30, 000 more efficient than prior work. Finally, in Section 3.3 we demonstrate via three behavioral studies that our model s predictions match human judgements on new, scaled-up tasks inaccessible to prior work. Section 3 is titled "Experiments" and includes "Qualitative analysis" and "Quantitative analysis" sections with tables and comparisons. |
| Researcher Affiliation | Academia | Kartik Chandra* MIT CSAIL Tony Chen* MIT Brain and Cognitive Sciences Tzu-Mao Li UC San Diego Jonathan Ragan-Kelley MIT CSAIL Joshua Tenenbaum MIT Brain and Cognitive Sciences |
| Pseudocode | Yes | Algorithm 1 Rejection sampling, as in prior work. Compare to our proposed method, Algorithm 2. Algorithm 2 Our bidirectional likelihood sampler Algorithm 3 Grow the bidirectional path tracer s cache (to be called repeatedly) |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or a link to a code repository for their method. It mentions |
| Open Datasets | No | The paper uses custom-designed domains/environments such as "Simple gridworld", "Doors, keys, and gems", "Word blocks", "food trucks domain", and a "multi-agent domain". While some are re-implementations or inspired by cited work, there is no explicit information (link, DOI, repository, or formal citation for dataset access) indicating that these specific datasets/environments are publicly available. |
| Dataset Splits | No | The paper discusses evaluating its algorithm using a specific number of Monte Carlo samples (e.g., 10, 1000, 10000 samples) and trials, and compares against a 'converged posterior'. However, it does not specify train, validation, or test dataset splits in the context of training a machine learning model or for data partitioning. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as CPU or GPU models, memory, or cloud computing specifications. It mentions using a "pre-trained Proximal Policy Optimization (PPO) controller" which implies computational resources were used, but no specifics are given. |
| Software Dependencies | No | The paper mentions using "stable-baselines3" and "Proximal Policy Optimization (PPO)" but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | Algorithm 2 lists specific parameters for the algorithm, such as "α, the strength of importance sampling" and "d, an average termination depth for Russian roulette". Section 3.1 mentions using "10 samples" for qualitative analysis. |