search

3ddiehead / Predictive_MarI-O

Sethbling's MarI/O with added memory mod, world state predictive mod, network state feedback mod
0 stars 0 forks source link

Move to Beta 3.0 #9

Open 3ddiehead opened 2 years ago

3ddiehead commented 2 years ago

Beta 3.0 includes a predictive differential so that unexpected things get special attention from the neural net. After Beta 3.0 and the analytics code are written, then compare 1.0, 2.0, and 3.0 for different qualities over time.

3ddiehead commented 2 years ago

Must first resolve fitness and mutation in the 2.1 predictor (it is almost completely ineffective)

3ddiehead commented 2 years ago

Currently the predictor aims for accuracy against the present frame, and it is both rewarded and penalized based on this accuracy comparison. BUT, what if the fitness of the predictor is not based on accuracy and instead based on the EFFECTIVENESS of the DIFFERENCE between the prediction and the present frame. That is, intentional inaccuracy, as long as it helps the predictor, could be beneficial.

This may be achieved naturally through the evolutionary process while the predictor simply focuses on accuracy, systematic inaccuracy only being retained as a byproduct of its usefulness in relation to the PHYSICAL fitness realm. OR, the predictor could be made to always overpredict (only benefits from accuracy / not penalized for overprediction), and its only restriction or penalization is its detriment to physical fitness.