Would have been less suspenseful to mention CDT and UDT in the introduction.
Top of second half of page 8, you say a strategy-selecting agent doesn’t update its action. This confused me, since strategies can have different observations map onto different actions—probably better to say something like it doesn’t update its strategy based on observations that are already in the strategy.
You switch over to talking about “algorithms” a lot, without talking about what that is. Maybe something like “An agent’s decision algorithm is the abstract specification of how it maps observations onto actions.” You can use this to explain the notation A()=s; both sides of the equation are maps from observations to actions, the right side is just concrete rather than abstract.
When you say “the environment is an algorithm,” you could cash this out as “the environment has an abstract specification that maps outputs of a strategy onto outcomes.”
Minor nitpicks:
Would have been less suspenseful to mention CDT and UDT in the introduction.
Top of second half of page 8, you say a strategy-selecting agent doesn’t update its action. This confused me, since strategies can have different observations map onto different actions—probably better to say something like it doesn’t update its strategy based on observations that are already in the strategy.
You switch over to talking about “algorithms” a lot, without talking about what that is. Maybe something like “An agent’s decision algorithm is the abstract specification of how it maps observations onto actions.” You can use this to explain the notation A()=s; both sides of the equation are maps from observations to actions, the right side is just concrete rather than abstract.
When you say “the environment is an algorithm,” you could cash this out as “the environment has an abstract specification that maps outputs of a strategy onto outcomes.”