Oregon State University School of Electrical Engineering and Computer Science End-User Programming of Intelligent Learning Agents Prasad Tadepalli, Ron.

Oregon State UniversitySchool of Electrical Engineering and Computer Science

End-User Programming of Intelligent Learning Agents

Prasad Tadepalli, Ron Metoyer, and Margaret Burnett

In conjunction with the EUSES Consortium:

End Users Shaping Effective Software

2

Prasad Tadepalli:Machine Learning

• Scaling Average-reward Reinforcement Learning to large spaces

Relational Learning

• Relational learning from prior knowledge and sparse user input

• Relational Reinforcement Learning

3

• NSF CAREER Award winner (2003).

• Complexities of animated content.– Creating characters for training.– Emphasis on usability and realism.

• Real-time simulation of evacuation dynamics for large crowds.

Ron Metoyer: Computer Graphics &

Animation

4

Margaret Burnett: Visual & End-User Programming

• Project director: EUSES Consortium (End Users Shaping Effective Software)

• An ITR project by Oregon State, Carnegie Mellon, Drexel, Nebraska, & Penn State.

• Principal architect: • Forms/3, FAR

end-user programming support.

• Co-architect: • Functions for Excel users (a Microsoft Research

project).

5

Motivation

• Task Training– Sports– Military

Boston Dynamics Inc.

Electronic Arts

Who creates the training content?

6

Current Approaches

• Joystick Control: – User does all (once, not reusable).

• Scripting Languages– User does all (reusable program).

• Programming by Demonstration – User and system share.

• Autonomous Agents – System does all.

7

Application:Quarterback Training

• QB’s can benefit from 3D training content• Coaches:

– Do not program or animate.– Need responsive, semi-intelligent agents that perform

football tasks.

• Agents:– Should get better over time.– Should do so with few examples.

• Agent behavior: – Must morph over time (different opponents).

8

End-User Programming by Demonstration

• Generalizing from demonstrations is still an active area of research:– Some viable approaches for particular

assumptions, but not a solved problem.

• Other systems allow demonstrating only reactive behaviors.– Not used to train people strategy.– Largely distinct from machine learning.

9

Our Approach to End-User Programming

• Our approach: demonstrate goals and strategies to achieve the goals.– Allows generalization and planning by agents.– Thus, suited to training:

• Agents can simulate both “good” characters for training (desirable strategies) ...

• and “bad” characters (strategies we know they employ).

10

Example

• Goal: Get the football to Character A.– Demonstration: Start state, goal state.– Research issue: “What is relevant”?

• Any trees are ignorable background.• Character A can be any character.• The football is a unique object.

Start: Goal:

11

• Strategy 1: Pass it directly.– Demonstration: Passing to A.– “What’s relevant” issues arise again.

• Strategy 2: Pass it to B who passes to A.– New issue: recursiveness. (Need to learn a general

strategy of “get it to someone who can get it to closer to A”.)

Example (cont.)

12

Machine Learning Challenges

• Learning must be on-line.

• Users can only give a few examples.

• Provide a predictable model of generalization.

• Must include support for debugging.

• Must allow safety checks.

• Expressive representation language.

13

Strategy Languages

• Some high-level languages exist to express strategies, e.g., Golog, CML.

• Our plan: simpler rule-based languages, suitable for learning.– Starting point: our previous work on a

decomposition-rule language:

IF Condition(s) and Goal(s)

Then Subgoals(s1,s2,..sn)

While invariant conditions hold.

14

Requirements of the Learning Algorithms

• Follow HCI findings: – User motivation, attention, trust.

• Need transparent generalization procedure, e.g., no neural nets.

• Treat user input as examples of high-level specification of strategy...– ...and fill in the details. – User “steers” agent behaviors to correct faulty

generalizations.

• Assertions to monitor behavior.– Provided, Inferred, and propagated.

15

Learning from Exercises

• Generate examples automatically by searching for successful plans.

• Bottom-up learning of skills.– Learn how to solve simple problems

first.– Compose known strategies for solving

subgoals to solve more complex goals.

16

Oops! That’s Not Right!

• Debugging by end-user programmers.– When the agents pick the right strategy but

it doesn’t work right.– When the agents pick the wrong strategies.

• These provide negative examples to the learning component.

17

How to Support Debugging?

• User/system collaboration.– User helps narrow the problem. – System revises its rules and runs them on the

example until the user is satisfied.

• Testing and Assertions– Used for quality control, but designed

specifically for end users.– Assertions will be used to rule out bad

generalizations.

18

Debugging (cont.)

• Draws from our previous work on end-user software development: – WYSIWYT testing, fault localization,

and assertions.– Surprise-Explain-Reward strategy:

• Empirically driven research.• Draws from psychology

to motivate desired behaviors viasurprises (to arouse curiosity).

19

Research Issues

• How to learn from a small number of examples?

• How to let the user “speak” his/her own language?

• How to motivate the users and earn their trust?

• How to facilitate debugging and maintenance in a natural way?

• How to make learning safe?

20

Summary: The Research Question

• Is it possible to empower end users...

• ...to program in evolving task-training environments...

• ...using machine learning and programming by demonstration?

21

(The End)

22

Leftovers

23

24

How to Support Debugging?

• User/system collaboration.

• Builds on our previous work: – Motivating, suggesting, and supporting...– ...end-user testing, end-user fault

localization, and end-user assertions.

25

Web Navigation (**possibly cut)

Navigation of the web to satisfy a goal:• Students trying to find an appropriate

school that match their interests and constraints.

• Shoppers looking for bargain purchases.• Traders searching for appropriate stocks

to buy and sell. In each case, the system should learn to retrieve the target information efficiently.

26

Debugging

• Negative examples are used to specialize over-general rules.

• Maintain confidences of rules based on their support among the training examples and suggest possible incorrect rules.

• Encourage users to enter assertions to correct errors.

• Verify assertions during rule evaluation and warn the user if they are not valid.

27

Agent Behavior Control

Joystick Controlled Autonomous

Scripting languages Autonomous but“teachable”

End-User Agents-program by interaction-generalize

28