Recurrent Gaussian Processes
Authors: César Lincoln Mattos, Zhenwen Dai, Andreas Damianou, Jeremy Forth, Guilherme Barreto, Neil Lawrence
ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we evaluate the performance of our RGP model in the tasks of nonlinear system identification and human motion modeling. ... The results are summarized in Tab. 1 and the obtained simulations are illustrated in Fig. 2. |
| Researcher Affiliation | Academia | 1,5Federal University of Cear a, Fortaleza, Cear a, Brazil 2,3,6University of Sheffield, Sheffield, UK 1cesarlincoln@terra.com.br 2,3{z.dai,andreas.damianou}@sheffield.ac.uk 4jforth@iweng.org 5gbarreto@ufc.br 6N.Lawrence@dcs.sheffield.ac.uk |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper provides links to YouTube videos demonstrating generated motions, but no concrete access (e.g., repository link, explicit release statement) to the source code of the methodology described in the paper. |
| Open Datasets | Yes | The first real dataset, named Actuator and described by Sj oberg et al. (1995) 1, consists of a hydraulic actuator that controls a robot arm, where the input is the size of the actuator s valve opening and the output is its oil pressure. ... 1Available in the Da ISy repository at http://www.iau.dtu.dk/nnbook/systems.html. ... The second dataset, named Drives and introduced by Wigren (2010), is comprised by a system with two electric motors that drive a pulley using a flexible belt. ... Dataset available in http://www.it.uu.se/research/publications/reports/ 2010-020/Nonlinear Data.zip as DATAPRBS.mat, with input u1 and output z1. ... The motion capture data from the CMU database2 was used to model walking and running motions. ... 2Available at http://mocap.cs.cmu.edu/. |
| Dataset Splits | No | In the case of the artificial dataset we choose L = Lu = 5 and generate 300 samples for training and 300 samples for testing... For the real datasets we use L = Lu = 10 and apply the first half of the data for training and the second one for testing. ... Training was performed with the trajectories 1 to 4 (walking) and 17 to 20 (running) from subject 35. The test set is comprised by the trajectories 5 to 8 (walking) and 21 to 24 (running) from the same subject. No mention of a separate validation split was found. |
| Hardware Specification | No | No specific hardware details (e.g., CPU, GPU models, memory) used for running experiments were mentioned in the paper. |
| Software Dependencies | No | We use the MLP implementation from the MATLAB Neural Network Toolbox with 1 hidden layer. (No version number for MATLAB or the toolbox is provided.) |
| Experiment Setup | Yes | In the case of the artificial dataset we choose L = Lu = 5... We compare our RGP model with 2 hidden layers, REVARB inference and 100 inducing inputs... For the real datasets we use L = Lu = 10... We use the MLP implementation from the MATLAB Neural Network Toolbox with 1 hidden layer. We also include experiments with the LSTM network, although the task itself probably does not require long term dependences. The original LSTM architecture by Hochreiter & Schmidhuber (1997) was chosen, with a network depth of 1 to 3 layers and the number of cells at each layer selected to be up to 2048. LSTM memory length was unlimited, and sequence length was chosen initially to be a multiple of the longest duration memory present in the data generative process, and reduced gradually. During experiments with varying LSTM network configurations, it became clear that it was possible in most cases to obtain convergence on the training sets, using a carefully chosen network model size and hyperparameters. Training was organized around batches, and achieved using a learning rate selected to fall slightly below loop instability, and it was incrementally reduced when instability re-appeared. A batch in this context is the concatenation of fixed length sub-sequences of the temporal data set. Neither gradient limits nor momentum were used. ... We evaluate a 2 hidden layer REVARB with 200 inducing inputs, the standard GP-NARX model and a 1 hidden layer MLP with 1000 hidden units. The orders are fixed at L = Lu = 20. ... We train a 1 hidden layer REVARB model with the RNN sequential recognition model (two hidden layer 500-200 units). |