Transfer across tasks
Figure 7 in Section 4.5 compares two policies trained on Demon Attack: one with zero latency and one with mixed latencies. Panels show Demon Attack, Asterix, Atlantis, and AirRaid, evaluated at fixed delays of 0, 2, and 4 raw frames. Each displayed mean uses 20 evaluation episodes, as documented in the manuscript. Returns use each task's native reward scale.
data.csv is the committed local plotting input. Its columns contain the
task, evaluation delay, zero-latency-trained mean return, and mixed-latency-trained
mean return. It does not contain checkpoint identifiers, the mixed-delay
training distribution, per-episode records, or uncertainty estimates.
At four frames, the mixed-latency policy has higher returns on Atlantis (5,010 versus 2,000) and AirRaid (178 versus 15), but lower return on Asterix (178 versus 212). Asterix has the opposite direction at two frames (262 versus 202). Demon Attack is the training-task comparison (3,510 versus 96 at four frames).
Assets and reproduction
fig.pdf and fig.png are the existing vector and 300-DPI raster assets.
fig_caption.tex supplies the manuscript caption; its numbering comes from
LaTeX. plot.py reads only the local CSV and uses the shared
plot_style.mplstyle and paper_colors.py: gray denotes zero-latency
training and purple denotes mixed-latency training.
The chart uses lines and markers; exact point values are reported in the CSV
and manuscript text.
From the paper repository root:
python3 figures/fig21_demon_mixed_cross_task/plot.py
latexmk -pdf arxiv.tex
latexmk -pdf iclr2027.tex