3 ms·
I believe that mario bros is deterministic and doesn't involve any randomness. Did you implement random start for fair evaluations to make sure that he found pa
by rubenfiszel 10y ago
I believe that mario bros is deterministic and doesn't involve any randomness.
Did you implement random start for fair evaluations to make sure that he found patterns instead of learning what to do by heart (which is a lot simpler).
Random evaluations is a chain of random input at the start to randomize the environment a bit.
- ehrenbrav79 10y agoNo - I actually didn't use a random start. But when Mario dies far enough along in a level, he starts from near that point rather than from all the way back at the beginning, which kind of serves that purpose. I agree that what's most interesting is truly learning to generalize rather than just learning a sequence of moves that happen to work...
- rubenfiszel 10y agoWell not really since the fact that he dies and respawn at that point is still deterministic. Cool project though
- Retric 10y agoIf you introduce artificial randomness your not teaching it to play Supper Mario. One option might be to include randomly pressing the wrong button with some low probability.
- ehrenbrav79 10y agoI think that's the epsilon parameter. That's the chance that, instead of playing the move Mario thinks has the highest value, he plays a random one instead. That in theory should help prevent him from falling into a local minimum. This probability starts at 100% at the beginning of training and over the course of 1 million steps, ramps down to 1%
- leereeves 10y agoThis might suggest it has learned by heart, not by pattern: In the example video, during the three lives on the second level, it does worse on the same parts of later lives. Of course, to draw any conclusions, we'd need a lot more than one run.