Step Efficiency & Redundancy

The problem, from first principles

Two agents can both solve a task, but one in 4 steps and the other in 40, repeating searches, re-reading the same file and retrying a failing command unchanged. The slow agent costs ten times more, takes ten times longer and is more likely to wander off. Evaluation therefore measures not only whether a task succeeded but how efficiently. Step efficiency compares the number of steps taken with the minimum known for the task (and is zero if the task failed, since an efficient failure is not a virtue). Redundancy counts actions that exactly repeat an earlier action, the signature of a stuck loop.

From theory to code

Implement step_efficiency and redundancy_rate.

Constraints

• step_efficiency(n_steps, n_optimal, succeeded) returns 0.0 if succeeded is false, else min(1.0, n_optimal / n_steps). n_steps >= 1.
• redundancy_rate(actions) takes a list of (name, args) pairs (args is a dict) and returns the fraction of actions that exactly repeat an earlier action (same name and equal args). Return 0.0 for an empty list.
• Key actions with (name, json.dumps(args, sort_keys=True)).

Hints