Multi-Brain HyperNEAT Experiments in the Lone Patrol domain
Multi-Brain HyperNEAT
is a software framework that extends
HyperNEAT
to evolve agents that possess multiple brains. Different brains can
be used in different situations, making it easy to evolve multimodal
behavior.
Because agents must cooperate to solve the
the team patrol domain, it conflates the challenges
of multiagent coordination and multimodal
behavior. Thus, the lone patrol domain is
introduced to isolate
the multimodal aspect of team patrol.
This goal is accomplished by placing
only a single robot
in the same environment.
This robot is responsible for visiting
all branches of the plus sign.
To add to the domain's challenge,
the robot must visit the branches
in an order requiring the central
four-way intersection to be
handled in three different ways: turning left,
going straight, and turning right.
The videos on this page can be viewed in a playlist here.
Lone Patrol: One Module (1M)
This robot with one module visits way points 0 and 1 with no problem.
In order to turn right to reach way point 2, the robot instead turns
left several times, which slows it down.
Lone Patrol: Situational Policy Geometry (SPG)
When situational policy geometry is used, a different control
network takes over after each way point is reached.
The same network is used to go from way point 0 to 1 as is used
to go from way point 2 back to the starting point.
Lone Patrol: Two Modules with preference neurons (2M)
This two-module robot mostly uses only one module,
indicated by the color red on the module usage timeline.
It performs ok until the last step of the patrol, when it
turns back to way point 0 instead of returning home.
Lone Patrol: Three Modules with preference neurons (3M)
This robot has three modules, but only uses two to successfully patrol the maze.
The green module is used a majority of the time, while the blue module primarily
helps the robot turn around after visiting way points at dead ends. The blue module
is also used briefly to help get around corners.
Lone Patrol: Module Mutation Duplicate (MM(D))
Module mutation duplicate creates many modules which the robot does not
actually use. Only three modules seem to be required to patrol the maze:
1) most navigation is handled by the blue module, particularly in straight
portions of the maze, 2) turning in general is handled by the green module,
though it is also used when passing through the center of the maze going
straight, 3) a beige module is used briefly whenever a dead end is encountered
to start turning around, before the green module takes over. This robot also
uses the inefficient trick used by 1M to reach way point 2: instead of turning
right, make a bigger turn to the left.
Lone Patrol: Module Mutation Previous (MM(P))
Module mutation previous also creates many un-used modules.
This particular robot visits all three way points, but gets stuck
returning to the starting point. The blue module is used to navigate
hallways at first, but whenever a dead end is reached, a red module turns
the robot around, after which a beige module is used to navigate hallways.
The blue module is used again to reach way point 1, but is not used after
that point because the beige module has taken over hallway navigation.
A white module also shows up near the end, but seems to be primarily
associated with being stuck.
Lone Patrol: Module Mutation Random (MM(R))
This robot produced by module mutation random has a large number of unused modules,
but it also uses several modules to intelligently patrol the maze. Most straight
hallways are handled by the red module. The initial left turn to reach way point 0
is accomplished with the help of a black module that is not used again. A beige module
is used to turn around in dead ends, and a purple module is used to pass straight through
the center of the maze without turning. However, in the latter half of evaluation,
the purple and beige modules are also used for other purposes that are less clear.
The robot uses these modules to deal with being stuck on a corner, but despite the
unusual module usage near the end, the robot manages to return to the starting point.
Lone Patrol: Multitask (MT)
This multitask robot has a different network that takes over after each
way point is reached. The same network is used to navigate from way point
0 to 1 as is used to return from waypoint 2 to home. Unlike with situational
policy geometry, there is no geometric relation between these networks, because
each one is produced by separate outputs in the CPPN genome. The robot has no
trouble visiting all way points and returning home.