New method for eligibility traces with clipping - seems to outperform the previous PPO-based method with history buffers.
2.5.1
Renamed flag for representation anticipation to just "anticipation", tuning, different image encoder.
2.5.0
New representation anticipation feature. Allows for training layers on feed back from higher layers without those higher layers being trained yet by assuming eventual alignment.
Feature can be disabled with:
`h.params.representation_anticipation = False`
Upside: Much faster converge for long sequence tasks. Downside: A bit slower. Most useful when there are many layers.
2.4.6
Fixed an issue with the early stopping detection
2.4.5
Fixed encoder initialization to perform better at spatial bridging (learn across spatial distances).
2.4.4
ReLU value function now has a separate dendrite count parameter.