Proposed Learning Benchmarks:

 

 


 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 


F2000 Simulator benchmark task (by: Alex Kleiner)

 

BACKGROUD:

            The proposed simulator server and related documentation can be downloaded here.

 

FIRST TASK: A single robot learns the skill shootGoal

 

LEARNING TASK: The agent learns this skill with all levels of the keeper (e.g. static, random, blockBall, interceptBall ).

 

PARAMETERS:

Input

DistToGoal, DistToKeeper, AngleToGoal, AngleToKeeper

Output

transVel, rotVel, kickBall

Success

Goal

Failure

ballLost (not in Goal)

 

  

 

 

 

 

 

SECOND TASK: A single robot learns to support an offender.

 

LEARNING TASK: The scenario could be learned by randomly placing the offender with ball and the supporter (the learning robot) in front of the goal. The supporting robot has to reach a position and orientation, which allows the offender to shoot a goal indirectly via the supporting robot. The offender will automatically shoot if receiving the "commSignalReady" from the supporter. If the reflecting ball hits the goal, the supporter will be rewarded.

 

PARAMETERS:

Input

AngleToGoal, DistToGoal, AngleToKeeper,

DistToKeeper, AngleToOffender

Output

commSignalReady

Success

Goal

Failure

ballLost (not in Goal and not at offender)

 

  

 

 

    

  

 

  

THIRD TASK: The offender selects from one of the learned skills (not necessarily within hierarchical RL)

 

LEARNIG TASK: The offender has the possibilities to either shoot directly or to turn towards the supporter until receiving the commSignalReady. The scenario could be learned by randomly plaicing a static goalkeeper into the goal.

 

PARAMETERS:

Input

DistToGoal, DistToKeeper, AngleToGoal, AngleToKeeper,

DistToSupporter, commSignalReady

Output

transVel, rotVel, kickBall

Success

Goal

Failure

ballLost and not in Goal

 

  

 

 

    

   

 

 

 

^ TOP


 

Rescue Simulation benchmark task (By: Mazda Ahmadi)

 

BACKGROUD:

 

In this benchmark task, there was a need for some pre-implemented agents. We developed some simple agents for

that reason and are now ready for download.

      

Random fire brigade:

http://ce.sharif.edu/~arian/learn-sig/Firebrigade-1.tar.gz

 

Fire brigade with simple targeting algorithm:

http://ce.sharif.edu/~arian/learn-sig/FireBrigade-2.tar.gz

(This fire brigade targeting algorithms is based on

1- Distance of the agent to the building

2- Building's fireyness

3- Number of on fire buildings near that building)

 

Police forces:

http://ce.sharif.edu/~arian/learn-sig/police-sig.tar.gz

 

SCENARIO:

 

The 1/100 of Kobe city map is used.

The test will run on a single initial configuration (i.e. same Shindopoly, Galpoly and gisini).

There is no civilian and no ambulance in the city and there are only five police forces and five fire brigades.

 

 

LEARNING TASK:

 

The aim of the agents is just to extinguish fires.

Since the main challenge in this task is learning to coordinate, the agents shouldn't use a coordination protocol.

 

You have control of just 'n' out of 5 fire brigades and 'm' out of 5 police forces.

 

There are different initial situations:

 

·        n = 2 and m = 2

     

            I - Other 2 fire brigades extinguish their nearest fiery building

                  and police forces move randomly.

II - Other 2 fire brigades uses a specific targeting algorithm

            for selecting fiery buildings and police forces move randomly.

                  III - Other 2 fire brigades and 2 police forces use a specific

                  learning algorithm

IV - Other 2 fire brigades and police forces behavior is one of

                  the known behaviors. The agents are given a total simulation

            to learn their teammates behavior and the second run counts.

 

 

·        n = 5 and m = 5

 

(the order is from easier to harder)

 

 

^ TOP