This repository contains an implementation of a framework for training and testing Hidden Parameter Markov Decision Processes (to appear at NIPS 2017, pre-print available here) and other RL benchmarks.

We introduce a new formulation of the Hidden Parameter Markov Decision Process (HiP-MDP), a framework for modeling families of related tasks using low-dimensional latent embeddings. We replace the original Gaussian Process-based model with a Bayesian Neural Network. Our new framework correctly models the joint uncertainty in the latent weights and the state space and has more scalable inference, thus expanding the scope the HiP-MDP to applications with higher dimensions and more complex dynamics.

• Example We demonstrate the training and testing of a HiP-MDP with embedded weights on the 2-D navigation domain (Grid) in toy_example.ipynb.
• Simulators: source code to run various control domains. In particular,
  • Acrobot The Acrobot, an inverted double pendulum, introduced by Richard Sutton (1996) and summarized in Reinforcement Learning: An Introduction.
  • Grid A toy 2-D navigation domain developed to illustrate the concept of transfer through a HiP-MDP.
  • HIV Determining effective treatment protocols for simulated patients with HIV tracking their physiological response to separate classes of treatments. First introduced as an RL domain by Damien Ernst, et al. (2006).
• Utilities: scripts used to build, combine and run the components of the HiP-MDP.
  • BNNs contains the class module and support functions to build, train and evaluate Bayesian Neural Networks with $\alpha$-divergence minimization.
  • Experience Replay contains the class to build and sample experience buffers, used for training neural networks with experience replay. Can be used to sample uniformly or in a prioritized fashion (after Schaul, et al. (2015)).
  • Priority Queue is a class module used to facilitate prioritized experience replay which implements a (Max) Binary Heap.
  • DQN is a class module that contains the code to build deep Q-networks. Currently defaulted to be a Double DQN (van Hasselt, et al. (2016)).

Python 2.7.12
tensorflow 0.12.1
numpy 1.11.1
autograd 1.1.7
seaborn 0.7.1

• Taylor Killian
• Samuel Daulton
• George Konidaris
• Finale Doshi-Velez

The source code and documentation are licensed under the terms of the MIT License.

• Harvard DTAK
• Harvard Paulson School of Engineering and Applied Sciences
• MIT Lincoln Laboratory Lincoln Scholars Program
• Alpha divergence bayesian neural network adapted from Jose Miguel Hernandez Lobato's original code
• HIV treatment and acrobot simulators adapted from RLPy's simulators
• Priority queue adapted from Kai Arulkumaran's Atari repository
• Deep Q Network follows Juliani's implementation