quantilization as a super low resolution discretization in the direction of (something like) geometric rationality.
level 1: take top 20% options of utility function and do a equiprobably random action
level 2: take top 40% options of utility function, give top 50% (the original top 20%) of those double weight and pick a random one
level 3: take top 60% of options, give top 40% double weight, give top 20% triple weight
level n+1: make smaller buckets to cover more of the range and have higher scoring buckets have bigger weights
beyond discrete: propability of picking an option is proportional to the its utility rank (“infinidesimally small buckets”)
quantilization as a super low resolution discretization in the direction of (something like) geometric rationality.
level 1: take top 20% options of utility function and do a equiprobably random action
level 2: take top 40% options of utility function, give top 50% (the original top 20%) of those double weight and pick a random one
level 3: take top 60% of options, give top 40% double weight, give top 20% triple weight
level n+1: make smaller buckets to cover more of the range and have higher scoring buckets have bigger weights
beyond discrete: propability of picking an option is proportional to the its utility rank (“infinidesimally small buckets”)