Collecting Results with Handlers

So far, optimize returned only the single best result. In many cases you want to see every result to watch progress over the optimization. A handler does this: an object you attach with a handlers= argument that observes every result an optimization produces.

It collects the full result objects — FunctionResults and GradientResults — rather than the summary optimize returns; see Working with Results.

A handler that collects everything

from ropt.simple import HistoryHandler, optimize

history = HistoryHandler()
result = optimize(config, x0, objective, handlers=[history])
print(len(history.results))   # every evaluation from this run

Compare this with the report callback from Follow the progress: report is called once per evaluation, for one run. A handler is more general — it keeps or reacts to results, and, unlike report, the same handler can be reused across several sequential calls to optimize, accumulating results from all of them.

Example: restarting from the best point

For instance, restart the same optimization from the best point the previous run found, while collecting every result from every restart in one handler:

x0 = initial_values
for _ in range(3):
    result = optimize(config, x0, objective, handlers=[history])
    assert result.results is not None
    x0 = result.results.variables   # restart from the best point found so far

print(f"collected {len(history.results)} results across all restarts")

Restarting needs nothing special from ropt: each call to optimize is independent, so result.results.variables — the best point a run found — is the start point for the next one. The runnable script is examples/simple/restart.py, which Restarting from the Best Point walks through.

Other built-in handlers

ropt ships a few ready-to-use handlers, all imported from ropt.simple:

• HistoryHandler — keeps every result, as used above.
• ResultsHandler — keeps only one result: the best seen so far (default), or the most recent.
• DataFrameHandler — collects results into a pandas or polars table.

See Result handlers for the full list, and how to write your own.