Epsilon-greedy exploration — contrasts in danger source with → R-max

Concrete Problems names epsilon-greedy and R-max together as standard exploration policies that "make no attempt to avoid dangerous situations" (concrete-problems, §"6 Safe Exploration", p. 14), but the two are blind to danger for opposite reasons. Epsilon-greedy takes a uniformly random action with fixed probability; it does not seek out unknown states, it simply fails to avoid them when chance selects one. R-max, by contrast, is built on optimism in the face of uncertainty — it treats every unvisited state-action pair as if it yields maximum possible reward, which is precisely what gives it its near-optimal learning guarantee. That same optimism actively steers the agent toward whatever is unexplored, including states that are unexplored because they are catastrophic. R-max's efficiency guarantee and its exploration hazard come from the same mechanism, whereas epsilon-greedy's hazard is incidental to its lack of one.