v0.2.0

• Features
  • Added multi-armed bandit endpoint. (#255)
    • Implemented epsilon-greedy. (#255)
    • Implemented epsilon-first. (#335)
    • Implemented UCB1. (#354)
    • Implemented UCB1-tuned. (#366)
  • Added support for the L-BFGS-B optimizer. (#296)
  • Added GPU implementation for q,p-EI and its gradient computation. (#219)
    • Speed up GPU functions by redesign of memory allocation. (#297)
• Changes
  • Split up old schemas.py file into schemas/ directory with several subfiles (#291)
  • Improved Dockerfile, reducing Docker-based install times substantially, https://hub.docker.com/u/yelpmoe/ (#332)
    • Created min_reqs docker container which is a snapshot of all MOE third-party requirements
    • Created latest, which tracks the latest MOE build
    • Started releasing docker containers for each tagged MOE release (currently just v0.1.0)
  • GradientDescentOptimization (C++) no longer has a separate next_points output (#186)
  • LogLikelihood evaluate at point list and latin hypercube search now return status dicts like every other optimizer (#189)
    • status dicts also a little more informative/standardized now
  • Update C++ autodoc tools to handle the new gpu directory (#353)
  • Added __version__ to moe/__init__.py (#353)
• Bugs
  • Throw exceptions (C++) if num_multistarts or num_random_samples is 0 (#345)
  • combined_example endpoint was not passing kwargs through so users could not change the default server (#356)
    • fix sometimes dropped general kwargs (#358)
  • mean_var_of_gp_from_historic_data was also not passing kwargs (#359)