Yuan Zhou

Yuanyi Zhong, Yuan Zhou, Jian Peng

The control variates (CV) method is widely used in policy gradient estimation to reduce the variance of the gradient estimators in practice. A control variate is applied by subtracting a baseline function from the state-action value estimates. Then the variance-reduced policy gradient presumably leads to higher learning efficiency. Recent research on control variates with deep neural net policies mainly focuses on scalar-valued baseline functions. The effect of vector-valued baselines is under-explored. This paper investigates variance reduction with coordinate-wise and layer-wise control variates constructed from vector-valued baselines for neural net policies. We present experimental evidence suggesting that lower variance can be obtained with such baselines than with the conventional scalar-valued baseline. We demonstrate how to equip the popular Proximal Policy Optimization (PPO) algorithm with these new control variates. We show that the resulting algorithm with proper regularization can achieve higher sample efficiency than scalar control variates in continuous control benchmarks.

Yuanyi Zhong, Yuan Zhou, Jian Peng

Reinforcement learning from self-play has recently reported many successes. Self-play, where the agents compete with themselves, is often used to generate training data for iterative policy improvement. In previous work, heuristic rules are designed to choose an opponent for the current learner. Typical rules include choosing the latest agent, the best agent, or a random historical agent. However, these rules may be inefficient in practice and sometimes do not guarantee convergence even in the simplest matrix games. In this paper, we propose a new algorithmic framework for competitive self-play reinforcement learning in two-player zero-sum games. We recognize the fact that the Nash equilibrium coincides with the saddle point of the stochastic payoff function, which motivates us to borrow ideas from classical saddle point optimization literature. Our method trains several agents simultaneously, and intelligently takes each other as opponent based on simple adversarial rules derived from a principled perturbation-based saddle optimization method. We prove theoretically that our algorithm converges to an approximate equilibrium with high probability in convex-concave games under standard assumptions. Beyond the theory, we further show the empirical superiority of our method over baseline methods relying on the aforementioned opponent-selection heuristics in matrix games, grid-world soccer, Gomoku, and simulated robot sumo, with neural net policy function approximators.

Yuan Zhou, Erin B. Wetherley, Paul D. Gader

In this paper, we model a pixel as a linear combination of endmembers sampled from probability distributions of Gaussian mixture models (GMM). The parameters of the GMM distributions are estimated using spectral libraries. Abundances are estimated based on the distribution parameters. The advantage of this algorithm is that the model size grows very slowly as a function of the library size. To validate this method, we used data collected by the AVIRIS sensor over the Santa Barbara region: two 16 m spatial resolution and two 4 m spatial resolution images. 64 validated regions of interest (ROI) (180 m by 180 m) were used to assess estimate accuracy. Ground truth was obtained using 1 m images leading to the following 6 classes: turfgrass, non-photosynthetic vegetation (NPV), paved, roof, soil, and tree. Spectral libraries were built by manually identifying and extracting pure spectra from both resolution images, resulting in 3,287 spectra at 16 m and 15,426 spectra at 4 m. We then unmixed ROIs of each resolution using the following unmixing algorithms: the set-based algorithms MESMA and AAM, and the distribution-based algorithms GMM, NCM, and BCM. The original libraries were used for the distribution-based algorithms whereas set-based methods required a sophisticated reduction method, resulting in reduced libraries of 61 spectra at 16 m and 95 spectra at 4 m. The results show that GMM performs best among the distribution-based methods, producing comparable accuracy to MESMA, and may be more robust across datasets.

Yuan Zhou, Anand Rangarajan, Paul Gader

The normal compositional model (NCM) has been extensively used in hyperspectral unmixing. However, most of the previous research has focused on estimation of endmembers and/or their variability. Also, little work has employed spatial information in NCM. In this paper, we show that NCM can be used for calculating the uncertainty of the estimated endmembers with spatial priors incorporated for better unmixing. This results in a spatial compositional model (SCM) which features (i) spatial priors that force neighboring abundances to be similar based on their pixel similarity and (ii) a posterior that is obtained from a likelihood model which does not assume pixel independence. The resulting algorithm turns out to be easy to implement and efficient to run. We compared SCM with current state-of-the-art algorithms on synthetic and real images. The results show that SCM can in the main provide more accurate endmembers and abundances. Moreover, the estimated uncertainty can serve as a prediction of endmember error under certain conditions.