Arrow Research search

Author name cluster

John Williams

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

5 papers
1 author row

Possible papers

5

AAAI Conference 2016 Conference Paper

Adaptable Regression Method for Ensemble Consensus Forecasting

  • John Williams
  • Peter Neilley
  • Joseph Koval
  • Jeff McDonald

Accurate weather forecasts enhance sustainability by facilitating decision making across a broad range of endeavors including public safety, transportation, energy generation and management, retail logistics, emergency preparedness, and many others. This paper presents a method for combining multiple scalar forecasts to obtain deterministic predictions that are generally more accurate than any of the constituents. Exponentially-weighted forecast bias estimates and error covariance matrices are formed at observation sites, aggregated spatially and temporally, and used to formulate a constrained, regularized least squares regression problem that may be solved using quadratic programming. The model is re-trained when new observations arrive, updating the forecast bias estimates and consensus combination weights to adapt to weather regime and input forecast model changes. The algorithm is illustrated for 0-72 hour temperature forecasts at over 1200 sites in the contiguous U. S. based on a 22-member forecast ensemble, and its performance over multiple seasons is compared to a state-ofthe-art ensemble-based forecasting system. In addition to weather forecasts, this approach to consensus may be useful for ensemble predictions of climate, wind energy, solar power, energy demand, and numerous other quantities.

PDF Details

NeurIPS Conference 1998 Conference Paper

Experimental Results on Learning Stochastic Memoryless Policies for Partially Observable Markov Decision Processes

  • John Williams
  • Satinder Singh

Partially Observable Markov Decision Processes (pO "MOPs) constitute an important class of reinforcement learning problems which present unique theoretical and computational difficulties. In the absence of the Markov property, popular reinforcement learning algorithms such as Q-Iearning may no longer be effective, and memory-based methods which remove partial observability via state-estimation are notoriously expensive. An alternative approach is to seek a stochastic memoryless policy which for each observation of the environment prescribes a probability distribution over available actions that maximizes the average reward per timestep. A reinforcement learning algorithm which learns a locally optimal stochastic memoryless policy has been proposed by Jaakkola, Singh and Jordan, but not empirically verified. We present a variation of this algorithm, discuss its implementation, and demonstrate its viability using four test problems.

PDF Details