Category talk:Log-Based Targeting

From Robowiki
Jump to navigation Jump to search

From old wiki's LogTargeting page

Targeting systems that utilize a detailed log of enemy data to predict where the target will be. The classic version of this is PatternMatching. Can work very well against simple bots and is currently a great way to hit WaveSurfers, but suffers from a long learning time and high processor loads, and the need for DealingWithSkippedTurns.

PatternMatching - Checking the targets current actions against things they had done in the past. The main goal is trying to balance precision with the fuzziness factor, and keeping it as fast as possible. Subsets of this are MogBot style, SymbolicPatternMatching, and whatever you want to call TronsGun.

NeuralTargeting - Using neural nets to predict enemy behavior. Rather obscure and very high level, but could have definite potential.

CommandReconstructionTargeting - A very obscure targeting method. Analyzes enemy behavior to calculate the commands that would give the result. Then once given the command list, attempts to reconstruct the movement code and use that to predict movement. The HolyGrail of Targeting if implemented.

SOOLTargeting - The log is self ordering. Locke uses it successfully. A description of the algorithm can be found on Locke's page.

Why are neural networks counted as log targeting? You use certain patterns (e.g. your enemy's velocity, heading, position) to train your neural network, but I doubt that you can find back what your enemy's bot was doing X ticks ago, like in a log. --Dummy

  • anyone??? I really don't think that NeuralNetworks fit the defenition of LogTargeting given above... -- Dummy
  • I can't say because I don't know very much about neural nets, but if you know a better category to put it in then I say go for it. --wcsv
  • If you look at bullet/item #3 on Albert's TrainingTheNN page, he mentions that training the network with ordered observations didn't work well for him. Instead, he logs his scans and trains the network with a random scan from history. Hence, the LogTargeting categorization. It's not the only way to train a network. Chomsky trains with observations as they come so he's not really using LogTargeting. --Corbos
  • I see. That means keeping a log isn't a requirement then. What do you guys think of 'Heuristics' as a category? Stuff like FuzzyLogic and GeneticAlgorithms should fit that category, I think. -- Dummy
  • Seems like a good idea to me. I think a lot of the stuff on the second level pages could use some review and/or reorganization. -- Voidious
There are no threads on this page yet.