Meta-Learning Acquisition Functions for Transfer Learning in Bayesian Optimization
Authors: Michael Volpp, Lukas P. Fröhlich, Kirsten Fischer, Andreas Doerr, Stefan Falkner, Frank Hutter, Christian Daniel
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present experiments on a simulation-to-real transfer task as well as on several synthetic functions and on two hyperparameter search problems. The results show that our algorithm (1) automatically identifies structural properties of objective functions from available source tasks or simulations, (2) performs favourably in settings with both scarse and abundant source data, and (3) falls back to the performance level of general AFs if no particular structure is present. |
| Researcher Affiliation | Collaboration | 1Bosch Center for Artificial Intelligence, Renningen, Germany 2ETH Zürich, Zürich, Switzerland 3Max Planck Institute for Intelligent Systems, Stuttgart/Tübingen, Germany 4University of Freiburg, Germany |
| Pseudocode | No | The paper describes the algorithm steps in text and provides a table (Table 1) translating Meta BO to RL framework, but no formal pseudocode or algorithm block is provided. |
| Open Source Code | Yes | To foster reproducibility, we provide a detailed exposition of the experimental settings in App. B and make the source code of Meta BO available online.3 https://github.com/boschresearch/Meta BO |
| Open Datasets | No | The paper mentions using datasets like 'Branin (D = 2), Goldstein-Price (D = 2), and Hartmann-3 (D = 3)' and 'precomputed results of training these models on 50 datasets' for HPO, but does not provide concrete access information (links, DOIs, formal citations with authors/year) for these datasets. |
| Dataset Splits | Yes | To determine when to stop the meta-training of Meta BO, we performed 7-fold cross validation on the training datasets. |
| Hardware Specification | No | For training Meta BO, we employed ten parallel CPU-workers to record the data batches and one GPU to perform the policy updates. Depending on the complexity of the objective function evaluations, training a neural AF for a given function class took between approximately 30 min and 10 h on this moderately complex architecture. This description lacks specific hardware model numbers or types (e.g., specific CPU or GPU models). |
| Software Dependencies | No | We used the implementation GPy (GPy, 2012) with squared-exponential kernels (Matern-5/2 kernels for the corresponding experiments on general function classes) with automatic relevance determination and a Gaussian noise model. We use the trust-region policy gradient method Proximal Policy Optimization (PPO) (Schulman et al., 2017) as the algorithm to train the neural AF. The paper mentions software by name but does not provide specific version numbers for reproducibility. |
| Experiment Setup | Yes | Table 4: Parameters of the Meta BO framework used in our experiments. Description Value in experiments BO/AF parameters Cardinality NMS of multistart grid Branin, Goldstein-Price 1000 Hartmann-3 2000 Simulation-to-real 10000 GPs (D = 1, 2, 3, 4, 5) 500, 1000, 2000, 3000, 4000 Cardinality NLS of local search grid NMS Number k of multistarts 5 Meta BO parameters Cardinality of ξglobal NMS Cardinality of ξlocal,t k Neural AF architecture 200 200 200 200, relu activations PPO parameters (Schulman et al., 2017) Batch size 1200 Number of epochs 4 Number of minibatches 20 Adam learning rate 1 10 4 CPI-loss clipping parameter 0.15 Value network architecture 200 200 200 200, relu activations Value coefficient in loss function 1.0 Entropy coefficient in loss function 0.01 Discount factor γ 0.98 GAE-λ (Schulman et al., 2015) 0.98 |