Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
207 lines (140 sloc) 7.79 KB

Language and behavior learning

Learning to follow instructions, interpret goals, model environment dynamics and select features in interactive environments.

Reinforcement Learning for Mapping Instructions to Actions.

Branavan, Chen, Zettlemoyer and Barzilay. ACL 09.

The original. Forms a joint MDP over the underlying environment and accompanying command, with the state space augmented to include the span of the command currently being processed and the action space augmented to move this span around. (This is an early version of hard attention!) Good results on a Windows troubleshooting task (apparently a hellish engineering effort) with carefully shaped rewards, as well as a block game with sparse rewards (for which I have a Scala reimplementation in

Learning to Interpret Natural Language Navigation Instructions from Observations.

Chen and Mooney. AAAI 11.

This is "supervised" in the sense that training commands are accompanied by demonstrations in the low-level action space. But there are many utterances with different meanings that might be consistent with a particular sequence of low-level actions ("go to the chair" vs "go to the red hallway" when the chair is located in the red hallway---the commands have the same extension but different intensions). Here we resolve the issue with a pipeline approach: first align words and phrases to actions and observations with which they are consistent across multiple paths, then use the lexicon to assign a specific high-level plan to each labeled low-level action sequence, and finally train a UT-style semantic parser on (command, high-level plan) pairs.

[N.B. People should stop using the SAIL dataset---it's tiny, existing work uses multiple versions of the data and multiple validation folds, lots of work just reports cross-validation rather than a held-out test set. See also note on Mei et al. 11 about evaluation metrics.]

Weakly Supervised Learning of Semantic Parsers for Mapping Instructions to Actions.

Artzi and Zettlemoyer. TACL 13.

As in Chen and Mooney 11, but with a more sophisticated logical language for describing plans and a proper latent-variable semantic parser à la Liang et al 11 (but Luke, so CCG). Can learn from both demonstrations and sparse reward on successful execution.

Listen, Attend, and Walk: Neural Mapping of Navigational Instructions to Action Sequences.

Mei, Bansal and Walter. AAAI 16.

First simple neural instruction follower for low-level actions: learns an instruction follower with seq2seq+attention and gets pretty good performance. The dataset is tiny by neural standards, so kind of surprising that this works at all.

[Results are not actually SOTA, as claimed in the paper---previous work starts with the agent facing in a fixed absolute direction, but this paper starts with the agent facing in a fixed direction relative to the correct move, which makes it trivial to guess the first action. Scores are artificially inflated as a result.]

Unified Pragmatic Models for Generating and Following Instructions

Fried, Andreas and Klein. NAACL 18.

Train a Mei et al.-style RNN model as well as an inverse model for generating instructions given actions. Ensembling these together corresponds to an implementation of a Rational Speech Acts model (Frank & Goodman 12), and gives good performance on both instruction following and generation. Not quite as good as semantic parsers on SAIL, but better on a bunch of other tasks.

Grounded Language Learning in a Simulated 3D World

Hermann, Hill, and a multitude.

Obvious: Mei et al without attention, but trained via RL instead of demonstrations using standard tricks. Fake language, fake environments, but hard vision; shows that you can scale simple instruction following models up if you have more compute than God.

Understanding Natural Language Commands for Robotic Navigation and Mobile Manipulation

Tellex, Kollar, Dickerson, Walter, Banerjee, Teller and Roy. AAAI 11.

First example I know of trying to predict a high-level cost function for consumption by a planner, rather than a low-level sequence of actions. Here the structure of a command is used to deterministically generate a command-specific CRF specifying a cost function that can be used for planning. The model is initially trained essentially discriminatively, to maximize the probability of correct groundings given a cost function and a command. These are predicted on held-out data with high accuracy, but this doesn't get us the ability to actually follow instructions. So there is also an "end-to-end" evaluation section in which groundings and cost functions are jointly inferred, and users asked if the resulting trajectories match the natural language descriptions. (This evaluation is only performed without any ground truth info for the 30 most confident test set predictions and works about half the time.)

Alignment-Based Compositional Semantics for Instruction Following

Andreas and Klein. EMNLP 15.

Similar in spirit to the Tellex et al. 11, but learns from demonstrations rather than cost functions by putting the low-level action sequence into the CRF. "Plans" at test time by doing inference over CRF variables.

Grounding English Commands to Reward Functions

MacGlashan et al. RSS 15

As in Andreas & Klein 15, but explicitly formulates the learning problem as IRL and uses a more sophisticated class of MDPs. (Note that IRL is really the same thing as CRF inference in both the maxent and max-margin formulations.) Toy experiments.

Learning to Win by Reading Manuals in a Monte Carlo Framework

Branavan, Silver and Barzilay. ACL 11.

This (2011!) paper is pretty wild: it learns a policy represented as a deep Q network (!) that scores actions based on the current environment state and soft attention (!) to dependency parses (!) extracted from a Civ II strategy guide. The Q network is trained from Monte-Carlo reward estimates rather than normal Q learning updates. Notably, we're just learning to play one game here; the "text" is really just a fixed blob of features that doesn't change during training time. So information-theoretically it shouldn't help at all---presumably what makes any of this work is that text-overlap features bias the learner toward particular strategies by correlating moves across turns.

Reading to Learn: Constructing Features from Semantic Abstracts

Eisenstein, Clarke, Goldwasser and Roth. EMNLP 09.

Approximately as in Branavan et al. 11, aims to provide good high-level features to a learner by extracting them from textual side-information. Specifies a generative model of text given logical specifications of high-level features using a glorious Bayesian nonparametric model. Evaluated on strategy games by training a classifier to predict legal moves from an extremely small number of examples.

Deep Transfer in Reinforcement Learning by Language Grounding

Narasimhan, Barzilay and Jaakkola

Deep RL through a value iteration network, using side-information from text to predict representations of entities (from which it can be inferred how they move, what rewards they provide, etc.). Faster learning on the training task, and fast adaptation to held-out tasks.