I am experimenting with logging all research related thoughts, readings and questions here. This blog also served as a way to learn Jekyll.

# Pruning in Loopy BP with large domains

Written on September 24, 2016

# Rank of translation options

Written on September 23, 2016

Translations were generated with cube and stack size of 400 and translation options 20. The target phrases used in the best translation were traced back to the translation options list and the rank positions were plotted in a histogram.

# Ignorability and randomized trials

Written on September 20, 2016

We can explain our intuition for why we randomized control trials seem to express causality.

# Meta Actions in MDPs

Written on September 9, 2016

Often in planning the lowest level of actions are defined and we hope the learning algorithm will eventually learn to compose these low level actions in sensible ways. One way to speed up convergence of the various iteration algorithms (value iteration, policy iteration etc) is to define not just low level actions, but pre-set sequences of actions as well. Instead of only sampling low-level actions the algorithm can also pick a meta action, which then executes an entire sequence. We could also learn these meta actions automatically as this paper describes.

# Parameters, Estimators and Estimates

Written on September 8, 2016

The causal inference class today covered some very basic ideas that I frequently have problems with. I ‘know’ what estimates and estimators are but could not clearly explain them to a lay person (which probably reflects some holes in my knowledge). Today’s class cleared some basic ideas.

# Macaronic Ice Berg

Written on September 8, 2016

The last slide in my ACL talk gave a vague idea of all the possible future tasks that are open issues.

# Speed Vs. Accuracy in Moses

Written on September 4, 2016

I ran a few experiments for Speed Vs. Accuracy in Moses SMT. 3000 sentences were translated from de, fr, es to en, while varying the following params:

# Thoughts on using MTCS for macaronic learning

Written on September 2, 2016

Monte Carlo Tree Search (MTCS) is a commonly used technique to solve/simplify the search problem that arises during game playing. An agent must consider multiple moves which can recursively lead to more moves to consider. This exponential tree search is not feasible in practice. Book-keeping in MTCS involves keeping track of wins and losses possible from each game state. In MTCS method can be divided into the following steps:

# Refinements to macaronic user modeling

Written on August 27, 2016

Things to consider regarding the testing of macaronic user model performance.

# RL Notes Chapter 2

Written on August 26, 2016

Definitions:

# Test equations

Written on June 23, 2016

this is a test here are some latex equations. $a = b + c + \frac{1}{2}$

# Reading list and review items

Written on May 12, 2016

In no specific order: