Dynamic Programming Boosting for Discriminative Macro-Action Discovery
Authors: Leonidas Lefakis, Francois Fleuret
ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluation is presented for the proposed method on tasks where change-points arise naturally as part of a classification problem. Finally we show the applicability of the algorithm to macro-action discovery in imitation learning and demonstrate it allows us to solve complex image-based goal-planning problems with thousands of features. 4. Experiments In order to highlight the strengths of the proposed algorithm we present a series of experiments in three distinct settings. |
| Researcher Affiliation | Academia | Leonidas Lefakis1,2 LEONIDAS.LEFAKIS@IDIAP.CH Franc ois Fleuret1,2 FRANCOIS.FLEURET@IDIAP.CH 1Idiap Research Institute, Martigny, Switzerland 2 Ecole Polytechnique F ed erale de Lausanne, Lausanne, Switzerland |
| Pseudocode | Yes | Table 1. Alternating DPBoost procedure. q, f 1 q Boost 0, {(xn, yn)}N n=1 for k = 1, . . . , K 1 do (qk+1 1 , . . . , qk+1 N ) DPSeq f k 1 , . . . , f k Q, {(xn, yn)}N n=1 q, f k+1 q Boost f k q , {(xn, yn)}n: qk+1 n =q |
| Open Source Code | No | The paper does not contain any statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We run experiments using the TIMIT dataset which is a acoustic-phonetic corpus used in the development and evaluation of automatic speech recognition systems. |
| Dataset Splits | No | The paper mentions using training sets and performing test runs, but it does not provide specific details about dataset splits, such as exact percentages or sample counts for training, validation, and test sets. For instance, in Section 4.2 it states "We isolate from the training data, and for each user, utterances of each phoneme... We then create conversations..." and in 4.3.3 "train three DPBoost goal-planners" but lacks precise split ratios or counts. |
| Hardware Specification | No | The paper mentions the use of "OGRE 3D graphics rendering engine" and "BULLET physics engine" for the simulation environment in Section 4.3. However, it does not provide any specific details about the hardware (e.g., CPU model, GPU model, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions the use of "OGRE 3D graphics rendering engine" and the "BULLET physics engine" in Section 4.3, but it does not specify any version numbers for these or any other software components. |
| Experiment Setup | Yes | We set a bound of T = 20 which is twice more than the actual number we are expected to meet (Section 4.1). We set the number of macro-classes equal to the number of speakers (Section 4.2). We set the upper limit T on the number of macro-classes to be twice the true number of speakers (Section 4.2). We set the upper limit on the number of macro-actions to {2, 3, 4} respectively (Section 4.3.3). The possible actions that the avatar can take...move forward by 3cm, and turn left or turn right which alter the avatar s orientation by π/300 (Section 4.3.1). The data xt available to DPBoost consist of the RGB values of a scaled down version (from 320 240 to 64 48 pixels) of the avatar s current view (Section 4.3.3). |