TaskMet: Task-driven Metric Learning for Model Learning
Authors: Dishank Bansal, Ricky T. Q. Chen, Mustafa Mukadam, Brandon Amos
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach through experiments conducted in two main settings: 1) decision-focused model learning scenarios involving portfolio optimization and budget allocation, and 2) reinforcement learning in noisy environments with distracting states. The source code to reproduce our experiments is available here. |
| Researcher Affiliation | Industry | Dishank Bansal Ricky T. Q. Chen Mustafa Mukadam Brandon Amos Meta Work done as part of the Meta AI residency program. |
| Pseudocode | Yes | Algorithm 1 Task Met: Task-Driven Metric Learning for Model Learning |
| Open Source Code | Yes | The source code to reproduce our experiments is available here. |
| Open Datasets | Yes | We use three standard resource allocation tasks for comparing task-based learning methods [Shah et al., 2022, Wilder et al., 2019, Donti et al., 2017, Futoma et al., 2020]. These settings are implemented exactly as in Shah et al. [2022] and have task losses defined by Ltask(θ) := E(x,y) D[g(z (ˆy), y)] (8) where z (ˆy) := arg minz g(z, ˆy) and g(z, y ) is some combinatorial optimization objective over variable z parameterized by y . |
| Dataset Splits | No | The paper mentions evaluating methods on "10 different datasets, with 5 different seeds used for each dataset" and discusses "test decision quality" and "test prediction errors". However, it does not explicitly provide specific training, validation, or test split percentages or sample counts for any of the datasets used. |
| Hardware Specification | No | The paper does not provide specific hardware details such as CPU or GPU models, or cloud computing instance types used for running the experiments. |
| Software Dependencies | No | The paper mentions replicating experiments using existing codebases (e.g., from Shah et al. [2022]) and building upon their source code for the MBRL experiments. However, it does not list specific software dependencies with their version numbers (e.g., Python 3.x, PyTorch 1.x, specific solver versions). |
| Experiment Setup | Yes | Table 4: Hyper-parameters for Decision Oriented Learning Experiments. Table 5: Hyper-parameters for the Cart Pole experiments. These tables list specific values for learning rate, hidden layer sizes, warmup steps, inner iterations, implicit derivative batchsize, and implicit derivative solver. |