Next: Tunable Parameters
Up: Novelty Engine
Previous: Novelty Engine
With the main population/novelty engine setup, the idle time -- while the system
is patiently waiting for human games to come in3.3 -- is devoted to the fabrication of the best opponents we can come up with.
It is expected that an evolutionary algorithm should increase the quality of
the new entities with every each generation. If there was no feedback from the
front end back to the novelty engine, the latter would remain oblivious to the
information being collected by the selective process acting on the foreground.
We have proposed and implemented two ways for surviving agents to come back
and reproduce in the novelty engine.
- 1.
- Using them as trainers. A reasonable setting for coevolving Tron agents
by self-play alone would have f=0. By setting f=15, some of the
champions-against-people come back to the novelty engine to act as trainers.
The hope is that, by transference, new agents are favored which are themselves
good against people.
- 2.
- Reintroducing them in the population. Successful agents can be simply
reintroduced in the population to let them compete with the other coevolving
robots and reproduce, provided they are successful against their peers.
Next: Tunable Parameters
Up: Novelty Engine
Previous: Novelty Engine
Pablo Funes
2001-05-08