Learning to Plan in High Dimensions via Neural Exploration-Exploitation Trees
Authors: Binghong Chen, Bo Dai, Qinjie Lin, Guo Ye, Han Liu, Le Song
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct thorough experiments to show that NEXT accomplishes new planning problems with more compact search trees and significantly outperforms state-of-the-art methods on several benchmarks. In this section, we evaluate the proposed NEXT empirically on different planning tasks in a variety of environments. Comparing to the existing planning algorithms, NEXT achieves the state-of-the-art performances, in terms of both success rate and the quality of the found solutions. We further demonstrate the power of the proposed two components by the corresponding ablation study. We also include a case study on a real-world robot arm control problem at the end of the section. |
| Researcher Affiliation | Collaboration | Binghong Chen1, Bo Dai2, Qinjie Lin3, Guo Ye3, Han Liu3, Le Song1,4 1Georgia Institute of Technology 2Google Research, Brain Team 3Northwestern University 4Ant Financial |
| Pseudocode | Yes | Algorithm 1: Tree-based Sampling Algorithm; Algorithm 2: NEXT :: Expand(T = (V, E), U); Algorithm 3: Meta Self-Improving Learning |
| Open Source Code | No | The paper provides a link to code used for generating maze maps for one of the benchmark environments, but it does not provide concrete access to the source code for the main methodology (NEXT) described in the paper. |
| Open Datasets | No | The paper states that workspace maps and obstacles were 'randomly generated' or 'uniformly randomly', and for the case study, planning problems were 'randomly generated'. While code for generating maze maps is linked, the paper does not provide explicit access (link, DOI, specific repository, or citation to an established benchmark) to the *datasets themselves* for public download or use. |
| Dataset Splits | No | The paper mentions training on 'the first 2000 problems' and reserving 'the rest for testing' but does not explicitly specify a separate validation dataset split or methodology. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory, or cloud instance types) used to run its experiments. |
| Software Dependencies | No | The paper mentions tools like 'Open Rave simulator (Diankov, 2010)' and 'C++ OMPL ( Sucan et al., 2012) implementation', but it does not provide specific version numbers for ancillary software dependencies required for replication (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We set the maximal number of samples as 500 for all algorithms. The value of the annealing ϵ was set as the following: 1, if i < 1000, 0.5 0.1 (i 1000)/200 , if 1000 i < 2000, 0.1, otherwise, with i denoting the problem number. In our experiments, we set the values of the hyper-parameters to be (d, de, da, p) = (15, 64, 8, 8). |