Return to Dr. Jacob Schrum's Homepage

Multi-Brain HyperNEAT Experiments in the Team 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. The team patrol domain is divided into two tasks: advance, in which the three robots spread out to the three segments of the plus sign, and return, in which the robots must return to their original starting positions. The videos on this page can be viewed in a playlist here.

Team Patrol: One Module (1M)

Each robot in the team has a separate network controller, but they are all generated using the same one module CPPN using Multi-Agent HyperNEAT. An input signal in each controller network indicates when the robots should return from their patrol. All robots manage to reach their patrol way point, but one of them gets stuck when the return signal activates.

Team Patrol: Situational Policy Geometry (SPG)

Use of situational policy geometry means that each of the three robots can have two separate network brains. Each is created using a different situation input in the CPPN genome. As a result, the controller networks do not need a return signal input. Instead, at the point when the return signal would be issued, a separate network is used, which allows all robots to return home successfully.

Team Patrol: Two Modules with preference neurons (2M)

Each robot has two separate network controllers, and decides which one to use at each moment based on outputs of preference neurons within each controller. One robot uses its green network (going to top), another uses its red network (going right), and only one makes use of both networks (going left). The network on the right gets stuck while returning.

Team Patrol: Three Modules with preference neurons (3M)

Each of the three robots have three separate network controllers. All use the blue and green modules, though one uses its blue module so seldom and briefly that the color does not show up clearly on the timeline. One robot also makes brief use of its red module. Two of the three robots manage to return home once the return signal activates.

Team Patrol: Module Mutation Duplicate (MM(D))

This application of module mutation duplicate results in each robot having four network controllers, though each robot generally only uses two modules. One robot uses the green module almost exclusively, but thrashes back and forth between green and blue when turning the corner on the way home. The other two robots use the red module to stay in the dead ends while waiting for the return signal, but use their primary module most of the time otherwise (green in one case, and blue in another). Thrashing module use also occurs with these robots, one of which gets stuck alternating back and forth between red and blue modules before completely returning home.

Team Patrol: Module Mutation Previous (MM(P))

Each robot of this team produced by module mutation previous has many unused controller networks. In fact, there are even "networks" that lack all links, or that do not connect many of the inputs to the outputs of the robots, making them unusable. However, there are viable controllers, and each team member makes use of two. One robot makes distinct use of blue and beige controllers, while the other two stick primarily to one controller and make only thrashing usage of another. The robot that favors the blue module gets stuck and fails to return home.

Team Patrol: Module Mutation Random (MM(R))

Each robot has five controllers produced by module mutation random, though each only uses two modules in order to perfectly patrol the maze and return home. All robots make exclusive use of a red module in order to reach their individual way points. When the return signal activates, all robots briefly make use of a green module to turn away from the dead end. One of these robots continues to use the green module to get it all the way back home, but the others only make thrashing use of the green module after responding to the return signal.

Team Patrol: Multitask (MT)

These robots advance into the maze and return home synchronously, each with the use of two separate controllers produced by their multitask CPPN. Each robot has one controller for advancing to its way point that was produced by one set of CPPN outputs, while the controllers used for returning home were produced by a second set of CPPN outputs. The result is the perfect behavior in the movie above.

Last Updated: 3/26/2016