Thread history

So thinking this through now, the walls in BerryBots have some interesting implications for PIF, too. For instance, if a log you're replaying includes that bot hitting the wall, what do you do? Or if a replay would cause you to hit the wall? In Robocode you might throw those out - it's rare and kind of a flukey situation, since you mostly never want to hit the wall. But there's no real disincentive to hitting walls in BerryBots, so it could be super frequent. Maybe you instead replay until the wall hit, then do a sort of single tick style re-projection for similar situations from that point on. And you probably want to precisely predict the wall bounce physics too.

Maybe what would make sense would be "fuzzy" wall collision handling that validates the projection against what happened in the PIF log:

1. Detect and log when the enemy hits a wall as part of the PIF log.
2. If the log that is being replayed does not contain a wall collision at a similar time, throw out projection if it collides with the wall by more than a certain margin (give leeway for the projection barely glancing by it)
3. If the log that is being replayed does contain a wall collision, throw out the projection unless the projection also contains a wall collision at a similar time (or at least a near-collision)

The idea would be to filter projections based on an approximate presence/absence of a collision matching the PIF log, in order to ensure the scenario of that projection is sufficiently similar to what you'll be predicting.

Very cool ideas. I guess you might need to allow for more fuzziness depending on the wall layout / wall hit frequency. I still like the idea of re-projecting from time of wall hit, but I'd really have to just try both and compare because I just don't have much sense of which would work better.

It also makes me wonder about doing the same in Robocode PIF. You could compare snapshots along the projections to further narrow down which situations match most closely or throw out irrelevant projections. You wouldn't have as many wall hits to compare, but you could still compare wall hits and distances or other stuff.

I guess the problem is you mainly want to examine stuff that causes you to alter your path, which requires knowing the future unless you're talking about walls that don't move. But if a past move log brings the enemy far away from all other bots on the field, and projecting that log now causes them to run into a swarm of bots, that might not be tough to predict and easy to realize you should throw it away.

That reminds me of some things I was pondering during the development of Glacier. I'd say Glacier's enemy distance segments are among the better ones in the meleerumble, but one thought I had but never got around to trying, was the notion of "single tick" style prediction of the whole field of other bots simultaneously, to predict the trend of group interactions rather than merely predicting an individual one's future behavior from it's immediate surroundings. After all, one bot making an agressive movement in one corner of the field could have a chain reaction causing bots on the other end of the field to move a bit, but the usual targeting systems aren't prepared to consider that.

If one wanted to really get tricky with that, there are ways one could use that "single tick" prediction in movement too... but yeah...

I had similar thoughts while making demonicRage, but never realized the value till reading yours. I think it would make a big improvement against advanced bots.. An "easy" way to test a diluted version of your concept, would be to predict the weekest bot on the field normally, then predict next weekest bot using the weaker bots predicted location and data... then predict 3rd bot using the 2 weaker bots predicted data..and so forth, so that the strongest opponent is predicted using all the predicted data of all the other bots.... That diluted method would likely be quick to jimmy up..

• edit* on second thought that breaks the 'bots interaction' part of your concept :) oh-well :P

Instead of classifying data using a single opponent at a time, classify by all opponents at the same time.

When predicting opponent´s A behaviour, instead of using only opponent´s A distance and velocity as input, use distance and velocity from other opponents too. The same principle applying to any kind of classification.

I thought about this before, but didn´t know how to deal with eliminated opponents. Thinking again, now I have some ideas.

One thing I've tried is attributes based on the force coming from an Anti-Gravity calculation. I thought it was going to be a killer feature, but despite being fairly rigorous to make sure it was doing what I wanted, I never got a performance gain out of it.

It seems like there must be a way to leverage that data, though.

I'd say "use distance from other opponents too" and "force coming from an Anti-Gravity calculation" are similar to what I used in Glacier's targeting actually. I used several dimensions which (more or less) measured the closest distance to another bot at different angle ranges.

With anti-grav systems, different ones will weight things differently, but in all cases the most prominant influences are the closest ones. As such, my goal was to create a measure with a small finite number of dimensions, that characterized the most prominant influences meaningfully.

In any case though, all those methods don't really consider the field-wide interactions that can occur, where movements can cause a chain reaction. For that, I doubt there is much of a reasonable option besides some sort of iterative process (maybe not per-tick, could be larger steps or iterations that are not time-based, but yeah).