Outline

A taste of in-progress works

Learning to shoot (finished) Applying the learned behavior to Multi-Agent scenarios

(Simulator Switch)

Learning to defend Learning to pass

Learning to Shoot a Moving Ball

Simulates real robots Learning is done using a Neural Network Generalizes to different situations Learned with 95% success video

Chaining Shots as Passes

Collaborative

Defending Against Aimed Shots

Adversarial Learn to block shots that are aimed at different parts of the goal

A Higher Level Multiagent Behavior

Shooting and Defending are low-level behaviors used in a multi-agent scenario. Soccer also requires higher-level multi-agent learning. Passing explicitly requires collaboration.
Passer decides to pass. Receiver prepares to receive.
Communication is necessary. Communication can be the result of Planning.

Decision Tree Learning

Learn the chances of a given pass succeeding

1. Choose a receiver.
2. Record the input data for that receiver.
3. Pass the ball.
4. Record the result.

174 continuous attributes. many values may be missing. trained on random receiver choices.

Pass will fail if > 0 else keep branching

C4.5 [release 8] decision tree interpreter      Tue Mar 19 13:49:09 1996
------------------------------------------
Decision Tree:

passer opponents ang6 > 0 : F (420.0/131.1)
passer opponents ang6 <= 0 :
|   passer receiver distance <= 18.5 :
|   |   passer teammates dist4 ang8 <= 1 :
|   |   |   passer teammates dist12 ang8 <= 1 : S (320.0/73.7)
|   |   |   passer teammates dist12 ang8 > 1 :
|   |   |   |   passer opponent3 distance <= 17.6 : F (5.0/2.3)
|   |   |   |   passer opponent3 distance > 17.6 : S (14.0/1.3)
|   |   passer teammates dist4 ang8 > 1 :
|   |   |   passer opponents dist4 ang4 <= 0 : S (12.0/2.5)
|   |   |   passer opponents dist4 ang4 > 0 :
|   |   |   |   passer opponent1 distance <= 19.1 : F (2.0/1.0)
|   |   |   |   passer opponent1 distance > 19.1 : M (2.0/1.0)
|   passer receiver distance > 18.5 :
|   |   passer opponent1 angle <= 14 :
|   |   |   passer opponents dist12 ang4 <= 0 :
|   |   |   |   receiver players dist12 ang8 > 1 : F (16.7/7.6)
|   |   |   |   receiver players dist12 ang8 <= 1 :
|   |   |   |   |   passer opponent2 angle <= 22 : F (258.3/106.9)
|   |   |   |   |   passer opponent2 angle > 22 : S (65.1/32.8)
|   |   |   passer opponents dist12 ang4 > 0 :
|   |   |   |   receiver opponent3 distance <= 6.9 : F (25.6/13.8)
|   |   |   |   receiver opponent3 distance > 6.9 : S (69.4/32.0)
|   |   passer opponent1 angle > 14 :
|   |   |   receiver teammate1 distance <= 19.7 :
|   |   |   |   passer opponent1 angle <= 22 : S (85.0/28.9)
|   |   |   |   passer opponent1 angle > 22 :
|   |   |   |   |   receiver players dist8 ang12 > 1 : M (2.0/1.0)
|   |   |   |   |   receiver players dist8 ang12 <= 1 :
|   |   |   |   |   |   passer opponent2 distance <= 23.1 : S (46.6/11.8)
|   |   |   |   |   |   passer opponent2 distance > 23.1 : M (9.1/4.6)
|   |   |   receiver teammate1 distance > 19.7 :
|   |   |   |   receiver opponent2 distance <= 15.6 : S (36.7/23.0)
|   |   |   |   receiver opponent2 distance > 15.6 : F (36.7/21.2)

Summary

Soccer is a great domain for multiagent learning. We learn both Collaborative and Adversarial behaviors. Learning is appropriate for both single-agent behaviors (shooting and defending) and multiagent ones (passing).

Levels of Learning

1. Shooting and Defending
2. Passing
3. Planning
   
4. Full Games

• About this document ...