This isolated variant of Ms. Pac-Man involves a pill eating task, in which there are no power pills, and a ghost eating task, in which Ms. Pac-Man spawns at each of the power pill points in each of the four mazes (devoid of pills) and has the usual limited time to eat the ghosts. These ghosts are spawned at locations experienced in the full blended game by a skilled agent. The videos show agents evolved using Modular Multiobjective Neuro-Evolution of Augmenting Topologies (MM-NEAT). Modular networks are superior, because evolution can discover unintuitive task divisions using preference neurons.
One Module: Typical Poor Behavior
Without power pills, the pill task is very hard, and most networks with only one module perform poorly in the pill eating task, as shown here. However, even poor One Module networks can do the ghost eating task fairly well because it is so easy, as is seen at the end of this video.
One Module: Best Outcome
Though most One Module networks do poorly in the pill eating task, this example shows that the best One Module networks are capable of eating all pills in the pill eating task, and of eating a good number of ghosts in the ghost eating task.
Two Modules: Escape Module
The most skilled modular approaches dedicate a seldom-used module to escaping dangerous situations in the pill eating task. Whenever this two module network uses its "escape" module, a green path is drawn on the map. This module is not used much early on, but in harder mazes, this module proves vital for getting certain pills. Whwnever a path is not drawn, this network is using its one other module. This module is used both a majority of the time in the pill task, and all of the time in the ghost task (which the agent performs well), but the seldom used escape module is still vital to the network's success.
Three Modules: Module For Each Task
Most modular networks only use two modules, like this three module network that dedicates one module to each isolated task: one for pill eating, one for ghost eating, and one that is not used at all. This module usage pattern is common, but not as effective as learning an escape module. In fact, this network gets trapped in the third level of the pill eating task because it made a bad decision when nearly surrounded by the ghosts. The two modules used by this network are identified by blue and red paths, and the video shows that only the blue path appears in the pill task, while only the red path appears in the ghost task.
Three Modules: Pill, Ghost, and Escape Modules
Though usage of only two modules is the most common behavior for modular networks, even among networks with more than two modules, a very small number of networks actually discover a way to intelligently use three modules. This network uses a blue module for the majority of the pill eating task, but starting in the second maze, brief usage of a green module for escaping can be seen. The network dies in the third level of the pill eating task, but it should be noted that this death occurs due to a random ghost reversal ... the network would have survived otherwise. Next, in the ghost eating task, a red module is used exclusively.
Module Mutation Duplicate: Pill, Ghost, and Escape Modules
This Module Mutation result is another rare case where a network uses three modules well. Module Mutation Duplicate gradually copies existing modules, then gives evolution a chance to modify the modules into distinct behaviors. This network has a green module for eating pills, a blue module for escaping in the pill task (plays an important role in third maze), and a red module for eating ghosts in the ghost task. This network manages to eat all pills and does a good job eating ghosts.
Multitask: Human-specified Task Division
Multitask networks have multiple modules, but use them according to a human-specified task division. That division dictates that a separate module is used in the pill task than is used in the ghost task, and this distinction is depicted with the use of green and blue paths. This is the best Multitask network evolved, and it does quite well, but even it dies in the pill task in the last maze. In general, this type of task division is not as good as the one using an escape module that evolution can discover with preference neurons.