|
NERO is a game of training and battle.
These movies are recorded from both parts of the game.
The behavior in these movies was discovered by rtNEAT evolution in real time; none was scripted.
Novel strategies and behaviors emerge on their own to fit the requirements of each task.
NEW MOVIES (Winter 2004-2005)
Training: Learn to approach the enemy. [27.4MB AVI]
In this learning scenario, the agents start out with empty brains. In the slider control panel in the lower right of the screen, the player assigns the agents the task of approaching the enemy. In a relatively short amount of time, the agents learn to approach the enemy on their own, without explicit code telling them how to accomplish that goal.
Training switch: Learn to approach the enemy, then learn to avoid the enemy. [14.7MB AVI]
In this scenario, the robots have already been trained to approach and attack the enemy. The
player then inhabits the enemy in first person mode and begins fighting back. The robots respond
by gradually appraching the player in fewer and fewer numbers.
Battle between simple aggressive team and simple avoiding team. [11.1MB AVI]
NERO battle scenarios do not involve active learning. Instead, the teams of agents have been trained previously and bring their trained behavior to the battlefield. This battle involves teams that have been trained in a straightforward way to either approach the enemy or avoid the enemy.
Battle between more complex teams. [12.9MB AVI]
In this battle scenario, the two teams were trained using more sophisticated strategies which produce more complex behavior in battle. The cooperative behavior evident in the blue team is not the result of scripting or an explicit cooperative output in the agents' networks. It is emergent behavior.
Note: The following two highly compressed movies require downloading and installing the DivX codec so that they can run on your system.
Download DivX here.
Battle between sophisticated teams [5.5MB AVI]
These teams learned on their own to cluster together in groups and keep a safe distance, as well as
pivot in and out of range. They display an effective and sophisticated battlefield strategy.
As with all NERO movies, this behavior was discovered by evolution and not scripted.
Complex Maze Dispersion Demo [1.2MB AVI]
The robots in this video were asked to weave their way through a complicated maze to
get to a flag at the other end. In addition, the "reward dispersion" slider was set to
high, meaning that they received more fitness if they did not all take the same route.
Computationally, this is a very difficult problem to solve: path-finding plus staying
away from other at the same time. However, rtNEAT (the technology behind NERO) was
able to evolve such a capability on its own with no path-panning algorithms built
into the system!
OLD MOVIES (Spring 2004)
Training
Note: In training, the trainees are blue and the enemy is red. Red smoke is from enemy shots,
blue is from the learners.
Training Mode: Learning to approach an enemy. [25.3MB AVI]
The NERO robots begin
the clip with empty brains and wandering aimlessly. Over a couple minutes, they learn to approach
the enemy. The learning can be seen in real-time. Notice how more and more robots exhibit the correct
behavior.
Training Mode: Training to avoid an enemy (first-person view). [6.4MB AVI]
This clip was taken in the middle of avoidance training. The robots are learning to avoid the player,
who is acting as their enemy. They can be observed running away.
Training Mode: Learning to navigate a maze. [6.4MB AVI]
The robots have been training inside a maze for several minutes before this clip
was recorded. At this point, the robots have learned maze navigation on their own without
any pathfinding algorithm.
Battle
Note: In battle one team is red and the other blue. Red smoke is from red team shots,
blue is from the blue team.
Battle Mode: Aggressive vs. Aggressive [5.5MB AVI]
Two teams that were trained to be aggressive are pitted against each other in battle.
Notice they both fearlessly meet in the middle. Small chase sequences can be observed
as robots break away from the main melee.
Battle Mode: Aggressive vs. Avoiding [4.6MB AVI]
An aggressive team takes on a team trained using avoidance training (see above clip of
avoidance training). Notice how easy it is to tell which is which. The avoidant team retreats
to a wall and eventually to a corner for a last stand.
Battle Mode: Team Firing Line (first-person view) [1.3MB MPEG]
The player takes first-person perspective during a battle and approaches a team that
has been trained to act like a firing line. Approaching a firing line is not very wise and
results in an unfortunate ending.
|
