en
0.25
0.5
0.75
1.25
1.5
1.75
2
Learning to Behave by Reading
Published on Aug 17, 20125944 Views
In this talk, I will address the problem of grounding linguistic analysis in control applications, such as game playing and robot navigation. We assume access to natural language documents that descri
Related categories
Chapter list
Learning to Behave by Reading00:00
Favorite Opening for my NLP Class00:11
"I´m sorry Dave, I´m afarid I can´t do that": Linguistics, Statistics and Natural Language Processing circa 2001 - Lilian Lee, Cornell University00:39
Can We Do It? - Objective / Challenge01:08
Semantic Interpretation: Traditional Approach01:43
Semantic Interpretation: Our Approach02:29
Learning from Control Feedback03:23
A Very Different View of Semantics04:14
Challenges05:28
General Setup06:39
Outline07:12
Mapping Instructions to Actions07:58
Learning Agenda08:45
Instruction Interpretation: Representation09:27
Learning Using Reward Signal: Challenges09:44
Reinforcement Learning: Representation10:14
Reward Signal11:33
Critique on the Critique13:34
Generating Possible Actions15:08
Model Parameterization15:40
Learning Algorithm16:02
Policy Function Factorization16:34
Example Features17:07
Windows Configuration Application17:36
Human Performance19:42
Results: Action Accuracy20:21
Applications of Instruction Mapping: WikiDo22:03
Active Learning: Performance23:39
Can We Do It? - Vision24:13
Outline: High-level strategy descriptions24:19
Solving Hard Decision Tasks (1)24:21
Solving Hard Decision Tasks (2)24:41
Case Study: Adversarial Planning Problem25:16
Research Agenda25:55
Leveraging Textual Advice: Challenges (1)26:56
Leveraging Textual Advice: Challenges (2)27:15
Leveraging Textual Advice: Challenges (3)27:26
Leveraging Textual Advice: Challenges (4)27:38
Leveraging Textual Advice: Challenges (5)27:58
Leveraging Textual Advice: Challenges (7)28:11
Model Overview: Monte-Carlo Search Framework28:24
Monte-Carlo Search (1)29:00
Monte-Carlo Search (2)29:29
Monte-Carlo Search (3)29:51
Model Overview: Our Algorithm30:27
Getting Advice from Text30:29
Modeling Requirements30:53
Sentence Relevance31:21
Predicate Structure31:53
Final Q function approximation32:11
Model Representation32:25
Learning from Game Feedback32:41
Parameter Estimation32:55
Experimental Domain33:04
Experimental Setup33:28
Results: Full Games34:10
Results: Sentence Relevance (1)35:11
Results: Sentence Relevance (2)36:40
Good Advice Helps!38:19
Monte-Carlo rollout38:55
Outline: General descriptions of world dynamics38:58
Solving Hard Planning Tasks39:27
Precondition/Effects Relationships39:32
How Text Can Help Planning40:17
Solution41:05
Learn Parameters Using Feedback from the World41:27
Experimental Domain41:36
Results41:38
Results: Text Analysis41:58
Conclusions41:59