kshitij judah eecs, osu dissertation proposal presentation

59
Developing Learning Systems that Interact and Learn from Human Teachers Kshitij Judah EECS, OSU Dissertation Proposal Presentation

Upload: theresa-jordan

Post on 28-Dec-2015

218 views

Category:

Documents


0 download

TRANSCRIPT

Developing Learning Systems that Interact and Learn from Human Teachers

Kshitij JudahEECS, OSU

Dissertation Proposal Presentation

Outline PART I: User-Initiated Learning

PART II: RL via Practice and Critique Advice

PART III: Proposed future directions for the PhD program Extending RL via practice and critique advice Active Learning for sequential decision making

PART IUser-Initiated Learning

• All of CALO’s learning components can perform Learning In The Wild (LITW)

• But the learning tasks are all pre-defined by CALO’s engineers:– What to learn– What information is relevant for learning– How to acquire training examples– How to apply the learned knowledge

• UIL Goal: Make it possible for the user to define new learning tasks after the system is deployed

User-Initiated Learning (UIL)

TIMELINE

Scientist

Sets sensitivity to confidential

Sends email to team

Sends email to a colleague

“Lunch today?”

Does not set sensitivity to confidential

Collaborates on aClassified project

Sends email to team

Forgets to set sensitivity to confidential

Research Team

Motivating Scenario:Forgetting to Set Sensitivity

“Please do not forget to set sensitivity when sending email”

Scientist

Research TeamTeaches CALO to learn to

predict whether user has forgot to set sensitivity

Sends email to team

CALO reminds user to set sensitivity

Motivating Scenario:Forgetting to Set Sensitivity

TIMELINE

SAT Based Reasoning System

SPARKProcedure

InstrumentedOutlook Event

s

user

Integrated Task Learning

user

Modify Procedure

Procedure Demonstration and Learning Task Creation

User Interface for Feature Guidance

User SelectedFeatures

user

Feature Guidance

Email + Related Objects

CALO Ontology

TrainedClassifier

Feature Guidance

Machine Learner

KnowledgeBase

Training Examples

Learning

Legal Features

SAT Based Reasoning System

Class Labels

User-CALO Interaction: TeachingCALO to Predict Sensitivity

Compose new email

SAT Based Reasoning System

SPARKProcedure

InstrumentedOutlook Event

s

user

Integrated Task Learning

user

Compose new email

Modify Procedure

Procedure Demonstration and Learning Task Creation

User Interface for Feature Guidance

User SelectedFeatures

user

Feature Guidance

Email + Related Objects

CALO Ontology

TrainedClassifier

Feature Guidance

Machine Learner

KnowledgeBase

Training Examples

Learning

Legal Features

SAT Based Reasoning System

Class Labels

User-CALO Interaction: TeachingCALO to Predict Sensitivity

SAT Based Reasoning System

SPARKProcedure

InstrumentedOutlook Event

s

user

Integrated Task Learning

user

Compose new email

Modify Procedure

Procedure Demonstration and Learning Task Creation

User Interface for Feature Guidance

User SelectedFeatures

user

Feature Guidance

Email + Related Objects

CALO Ontology

TrainedClassifier

Feature Guidance

Machine Learner

KnowledgeBase

Training Examples

Learning

Legal Features

SAT Based Reasoning System

Class Labels

User-CALO Interaction: TeachingCALO to Predict Sensitivity

SAT Based Reasoning System

SPARKProcedure

InstrumentedOutlook Event

s

user

Integrated Task Learning

user

Compose new email

Modify Procedure

Procedure Demonstration and Learning Task Creation

User Interface for Feature Guidance

User SelectedFeatures

user

Feature Guidance

Email + Related Objects

CALO Ontology

TrainedClassifier

Feature Guidance

Machine Learner

KnowledgeBase

Training Examples

Learning

Legal Features

SAT Based Reasoning System

Class Labels

User-CALO Interaction: TeachingCALO to Predict Sensitivity

Assisting the User: Reminding

• Logistic Regression is used as the core learning algorithm

• Features– Relational features extracted from ontology

• Incorporate User Advice on Features– Apply large prior variance on user selected features– Select prior variance on rest of the features through cross-

validation

• Automated Model Selection– Parameters: Prior variance on weights, classification threshold– Technique: Maximization of leave-one-out cross-validation

estimate of kappa ()

The Learning Component

• Problems:– Attachment Prediction– Importance Prediction

• Learning Configurations Compared:– No User Advice + Fixed Model Parameters– User Advice + Fixed Model Parameters– No User Advice + Automatic parameter Tuning– User Advice + Automatic parameter Tuning

• User Advice: 18 keywords in the body text for each problem

Empirical Evaluation

• Set of 340 emails obtained from a real desktop user

• 256 training set + 84 test set

• For each training set size, compute mean kappa () using test set to generate learning curves

• is a statistical measure of inter-rater agreement for discrete classes

• is a common evaluation metric in cases when the classes have a skewed distribution

Empirical Evaluation: Data Set

Attachment Prediction

Empirical Evaluation: Learning Curves

Importance Prediction

Empirical Evaluation: Learning Curves

• We intended to test the robustness of the system to bad advice

• Bad advice was generated as follows:– Use SVM based feature selection in WEKA to

produce a ranking of user provided keywords

– Replace top three words in the ranking with randomly selected words from the vocabulary

Empirical Evaluation: Robustness to Bad Advice

Attachment Prediction

Empirical Evaluation: Robustness to Bad Advice

Importance Prediction

Empirical Evaluation: Robustness to Bad Advice

• User interfaces should support rich instrumentation, automation, and intervention

• User interfaces should come with models of their behavior

• User advice is helpful but not critical

• Self-tuning learning algorithms are critical for success

Lessons Learned

PART IIReinforcement Learning via Practice

and Critique Advice

PROBLEM: Usually RL takes a long time to learn a good policy.

Reinforcement Learning (RL)

Teacher

behavior

advice

RESEARCH QUESTION: Can we make RL perform better with some outside help, such as critique/advice from teacher and how?

GOALS: Non-technical

users as teachers

Natural interaction methods

state

action

rewardEnvironment

Critique Data Loss L(θ,C)

Imagine: Our teacher is an Ideal Teacher (Provides All Good Actions)

Set of all Good

actions

Any action not in O(si) is suboptimal according to Ideal

Teacher

All actions

are equallygood

Advice Interface

Ideal Teacher

Advice Interface

Some good

actions

Some bad actions

Some actions

unlabeled

‘Any Label Learning’ (ALL)

Imagine: Our teacher is an Ideal Teacher (Provides All Good Actions)

Set of all Good

actions

Any action not in O(si) is suboptimal according to Ideal

Teacher

All actions

are equallygood

Advice Interface

Ideal Teacher

Learning Goal: Find a probabilistic policy, or classifier, that has a high probability of returning an action in O(s) when applied to s.

ALL Likelihood (LALL(,C)) :

Probability of selecting an action in O(Si) given state si

Critique Data Loss L(θ,C) Coming back to reality: Not All Teachers are Ideal !

and provide partial evidence about O(si)

Advice Interface

What about the naïve approach of treating as the true set O(si) ? Difficulties: When there are actions outside of that are equally good

compared to those in , the learning problem becomes even harder.

We want a principled way of handling the situation where either or can be empty.

Map 1 Map 2

Experimental Setup

Our Domain: Micro-management in tactical battles in the Real Time Strategy (RTS) game of Wargus.

5 friendly footmen against a group of 5 enemy footmen (Wargus AI).

Two battle maps, which differed only in the initial placement of the units.

Both maps had winning strategies for the friendly team and are of roughly the same difficulty.

Advice Interface

Difficulty:Fast pace and

multiple units acting in parallel

Our setup: Provide end-users

with an Advice Interface that allows to watch a battle and pause at any moment.

Goal is to evaluate two systems1. Supervised System = no practice session2. Combined System = includes practice and

critique

The user study involved 10 end-users 6 with CS background 4 no CS background

Each user trained both the supervised and combined systems

30 minutes total for supervised 60 minutes for combined due to additional time

for practice

Since repeated runs are not practical results are qualitative

To provide statistical results we first present simulated experiments

User Study

Simulated Experiments

After user study, selected the worst and best performing users on each map when training the combined system

Total Critique data: User#1: 36, User#2: 91, User#3: 115, User#4: 33.

For each user: divide critique data into 4 equal sized segments creating four data-sets per user containing 25%, 50%, 75%, and 100% of their respective critique data.

We provided the combined system with each of these data sets and allowed it to practice for 100 episodes. All results are averaged over 5 runs.

Simulated Experiments Results:Benefit of Critiques (User #1)

RL is unable to learn a winning policy (i.e. achieve a positive value).

Simulated Experiments Results:Benefit of Critiques (User #1)

With more critiques performance increases a little bit.

As the amount of critique data increases, the performance improves for a fixed number of practice episodes.

RL did not go past 12 health difference on any map even after 500 trajectories.

Simulated Experiments Results:Benefit of Critiques (User #1)

Simulated Experiments Results:Benefit of Practice (User #1)

Even with no practice, the critique data was sufficient to outperform RL.

RL did not go past 12 health difference.

Simulated Experiments Results:Benefit of Practice (User #1)

With more practice performance increases too.

Simulated Experiments Results:Benefit of Practice (User #1)

Our approach is able to leverage practice episodes in order to improve the effectiveness on a given amount of critique data.

Goal is to evaluate two systems1. Supervised System = no practice session2. Combined System = includes practice and

critique

The user study involved 10 end-users 6 with CS background 4 no CS background

Each user trained both the supervised and combined systems

30 minutes total for supervised 60 minutes for combined due to additional time

for practice

Results for Actual User Study

Results of User Study

-50 0 50 80 1000123456789

10

Health Difference Achieved by Users

SupervisedCombined

Health Difference

Num

ber o

f Use

rs

Results of User Study

Comparing to RL: 9 out of 10 users achieved 50 or more

performance using Supervised System 6 out of 10 users achieved 50 or more

performance using Combined System

Users effectively performed better than RL using either the Supervised

or Combined Systems.

RL did not go past 12 health difference on any map even after 500 trajectories.

-50 0 50 80 1000123456789

10

Health Difference Achieved by Users

SupervisedCombined

Health Difference

Num

ber o

f Use

rs

Results of User Study

Frustrating Problems for Users Large delay experience. (not an issue in many

realistic settings)

Policy returned after practice was sometimes poor, seemed to be ignoring advice. (perhaps practice sessions were too short)

Comparing Combined and Supervised: The end-users had slightly greater success

with the supervised system v/s the combined system.

More users were able to achieve performance levels of 50 and 80 using the supervised system.

-50 0 50 80 1000123456789

10

Health Difference Achieved by Users

SupervisedCombined

Health Difference

Num

ber o

f Use

rs

PART IIIFuture Directions

Future Direction 1: Extending RL via Practice and Critique Advice

Understanding the effects of user models: Study sensitivity of our algorithm to various settings of

model parameters. Robustness of our algorithm against inaccurate

parameter settings. Study benefits of using more elaborate user models.

Understanding the effects of mixing advice from multiple teachers: Pros: addresses incompleteness and lack of quality of

advice from a single teacher. Cons: introduces variations and more complex patterns

that are hard to generalize. Study benefits and harms of mixing advice.

Understanding the effects of advice types: Study the effects of feedback only versus mixed advice.

Future Direction 2: Active Learning for Sequential Decision Making

Current advice collection mechanism is very basic: An entire episode is played before the teacher. Teacher scans the episode to locate places where

critique is needed. Only one episode is played.

Problems with current advice collection mechanism: Teacher is fully responsible for locating places where

critique is needed. Scanning an entire episode is very cognitively

demanding. Good possibility of missing places where advice is

critical. Showing only one episode is a limitation, especially in

stochastic domains.

GOAL: Learner should itself discover places where it needs advice and query teacher at those places.

What about existing techniques?

Few techniques exist for the problem of ‘active learning’ in sequential decision making with an external teacher

All techniques make some assumptions that work only for certain applications

Some techniques request full demonstration from start state

Some techniques assume teacher is always available and request a single or multi-step demonstration when needed

Some techniques removes assumption of teacher being present at all times but they pause till the request for demonstration is satisfied

What about existing techniques?

We feel such assumptions are unnecessary in general

Providing demonstration is quite labor intensive and sometimes not even practical

We instead seek feedback and guidance on potential execution traces of our policy

Pausing and waiting for the teacher is also inefficient

We never want to pause but keep generating execution traces from our policy for teacher to critique later when he/she is available

What about Supervised Active Learning techniques?

Active learning is well developed for supervised setting

All instances come from single distribution of interest

Best instance is selected based on some criterion and queried for its label

In our setting, the distribution of interest is the distribution of states along the teacher's policy (or a good policy)

Asking queries about states that are far off of the teacher's policy is likely to not produce any useful feedback (losing states in Chess or Wargus)

Learner faces additional challenge to identify states that occur along the teacher's policy and query in those states

Proposed Approach for Supervised Setting

Ideally we should select sequences that directly maximize ECPL

Heuristic: Identify sequence that contains states with high probability of first disagreement (similar to uncertainty sampling)

Common Prefix Length

Unimportant States

States where first disagreement is likely

High ConfidenceExecution(most likely we are on right path)

Low ConfidenceExecution(most likely we are going to make a wrong turn)

(most likely we should not be here)

Proposed Approach for Supervised Setting

Fixed length sequences

Variable length sequences: compute optimal length that trades off benefit from critique versus cost of critique. We plan to use return on investment heuristic proposed by Haertel et. Al. (Return on Investment for Active Learning, NIPS Workshop on Cost Sensitive Learning, 2009)

Common Prefix Length

Unimportant States

States where first disagreement is likely

High ConfidenceExecution(most likely we are on right path)

Low ConfidenceExecution(most likely we are going to make a wrong turn)

(most likely we should not be here)

Proposed Approach for Practice Setting

Practice followed by active learning: let the agent practice, when critique session starts use heuristics from supervised setting to select best sequence for the current policy

Modify heuristic from supervised setting: Can we better discover sequences with first disagreement using information from practice

Based on self assessment during practice

Evaluation Plans

We will use RTS game Wargus as our domain

Extend our current tactical battle micromanagement scenarios with more and different types of combat units

Teach control policies for attack and maneuver behavior of individual units

Teach policies to control low-level actions in resource gathering tasks

User Studies: Conduct further user studies

Research Timeline

Period Research Topic

Aug 2010 Get proposal approved.

Sept-Oct 2010 Active learning in supervised setting, teacher models (for journal).

Nov-Dec 2010 Active learning in practice setting, advice patterns (for journal).

Jan 2011 Submit a paper to IJCAI/AAAI.

Feb-Mar 2011 Start writing journal paper, mixing advice (for journal).

Apr-May 2011 Finish writing journal paper, start writing dissertation.

June-Aug 2011 Finish writing dissertation, submit journal paper.

Sept 2011 defense.

Conclusion

Presented our completed work on: User-Initiated Learning RL via Practice and Critique Advice

Proposed future directions for the PhD program: Extending RL via practice and critique advice Active Learning for sequential decision making

Presented potential approaches and evaluation plans in order to carry out proposed work

Presented a timeline for completing the proposed work

Questions ?