msg7337 (view) |
Author: manuel |
Date: 2018-07-30.14:30:57 |
|
I close this issue because most of us agree with keeping the current behavior.
|
msg7314 (view) |
Author: silvan |
Date: 2018-07-24.12:40:04 |
|
I was interested in getting this issue from my radar :-)
I think the differences are too small to add a new option here, and thus also
vote for keeping the current behavior.
|
msg7313 (view) |
Author: jendrik |
Date: 2018-07-24.12:13:30 |
|
I also like Manuel's suggestion.
|
msg7312 (view) |
Author: malte |
Date: 2018-07-24.12:11:27 |
|
Every option has a cost in maintenance, documentation, understandability,
cognitive load and debuggability. I would rather not introduce a new one unless
there is a significant gain to be had from that. I like Manuel's suggestion.
What do the others think? (For example Silvan, you inquired about this recently,
so perhaps you're interested in the status?)
|
msg7311 (view) |
Author: mkatz |
Date: 2018-07-24.12:04:24 |
|
How about keeping both options, introducing yet another parameter to the search?
|
msg7310 (view) |
Author: manuel |
Date: 2018-07-24.11:47:39 |
|
Rewarding progress at the initial state does not show significant effects on
eager search.
For the experiment, I have set the boost value to 1000, because the standard
boost value is 0.
Configuration issue756-base is without reward at the initial state and
issue756-v2 is with a reward at the initial state.
http://ai.cs.unibas.ch/_tmp_files/heusner/issue756-v2-eager-issue756-base-issue756-v2-compare.html
I suggest keeping the current behavior of lazy search and eager search.
|
msg7301 (view) |
Author: silvan |
Date: 2018-07-12.14:17:42 |
|
Manuel, any news here?
|
msg7240 (view) |
Author: manuel |
Date: 2018-06-07.19:21:32 |
|
Eager search does not reward progress at the initial state in the current
version of Fast Downward. I suggested in a previous message to also test the
behavior of eager greedy on this issue. Let me do the experiments and report the
results here.
|
msg7235 (view) |
Author: malte |
Date: 2018-06-07.12:49:34 |
|
And all other things being equal, I agree Manuel: not rewarding the initial
state makes slightly more sense to me than rewarding it. But I think it should
be the same in lazy and eager search, and I did not check what currently happens
in eager search.
|
msg7234 (view) |
Author: malte |
Date: 2018-06-07.12:47:28 |
|
I looked at the original papers, and the descriptions are not formal enough to
be 100% sure how this case should be handled. For me, either rewarding or not
rewarding the initial state would make some sense, but the behaviour should be
consistent between lazy and eager search.
|
msg7228 (view) |
Author: silvan |
Date: 2018-06-07.10:09:23 |
|
So, again, regarding the issue, we should decide whether to keep the current
behavior or whether it makes more sense to switch to some more close to the
formal definition. Since I don't know about the latter, others who do should
speak up :-)
|
msg7219 (view) |
Author: mkatz |
Date: 2018-06-06.11:57:10 |
|
It looks like the current variant performs slightly better almost everywhere, which is
not what I observed in my experiments, but in my experiments there could be additional
factors. It might be worth taking another look if/when we decide to integrate red-black
heuristics.
|
msg7218 (view) |
Author: manuel |
Date: 2018-06-06.10:53:24 |
|
Revision issue756-base is without reward at initial state and issue756-v1 is
with reward.
You find the pull request on bitbucket:
https://bitbucket.org/manuel_h/downward/pull-requests/6/issue756/diff
|
msg7217 (view) |
Author: silvan |
Date: 2018-06-06.10:52:54 |
|
The differences do not seem to be very large. So I think the decision of which
variant to prefer boils down to which variant is closer to the definition of
progress and to what one would expect.
|
msg7216 (view) |
Author: silvan |
Date: 2018-06-06.10:44:43 |
|
Which version is which? Can you also please post a pull-request?
|
msg7215 (view) |
Author: manuel |
Date: 2018-06-06.09:26:16 |
|
The result of the experiment does not show significant differences over
different configurations and domains.
In my opinion we should not reward progress at the initial state, because all
definitions of progress that I have found in literature in context of boosting
preferred operators do define progress relative to heuristic estimates of
already expanded states. If the generation of the initial state is considered
as progress, it should have been defined explicitly.
Here you find the comparison tables:
http://ai.cs.unibas.ch/_tmp_files/heusner/issue756-v1-issue756-base-issue756-v1-compare.html#summary
|
msg7155 (view) |
Author: manuel |
Date: 2018-06-04.12:22:42 |
|
I will test and report the behavior of lazy search and ignore the other
observations for this issue. We may discuss the other issues offline.
|
msg7154 (view) |
Author: silvan |
Date: 2018-06-04.12:15:20 |
|
Then please change the reporting order as you think it is best.
|
msg7153 (view) |
Author: manuel |
Date: 2018-06-04.12:13:52 |
|
The reporting is consistent among search algorithms. I just feel that the flow
of reporting progress is broken by reporting the initial heuristic value and the
pruning method.
|
msg7152 (view) |
Author: silvan |
Date: 2018-06-04.12:09:06 |
|
I don't quite understand: is there a difference in the reporting of different
searches and you want to make it consistent? If yes, then this is good.
|
msg7151 (view) |
Author: manuel |
Date: 2018-06-04.12:06:35 |
|
I noticed that progress is also not rewarded in eager_greedy at the initial
state. I would test the behavior of rewarding progress at the initial state for
eager_greedy as well.
Moreover, I noticed that the initial heuristic values as well as the progress
methods are reported after reporting the first progress in eager and lazy
searches. I would expect both to be reported beforehand. This is a small issue
that could be fix within this issue.
What do you think?
|
msg6759 (view) |
Author: mkatz |
Date: 2017-12-20.12:27:40 |
|
Last before issue77 revision: 874838d3625e
After issue77 revision: bcb6cef0e11a
|
msg6758 (view) |
Author: malte |
Date: 2017-12-20.12:27:32 |
|
...or perhaps it's worth also testing cg and cea (both as single heuristics in
lazy search, of course with preferred operators), as we expected a behaviour
change that can be dramatic, but only for certain heuristics. Or at least that's
what Michael saw in his experiments.
|
msg6757 (view) |
Author: malte |
Date: 2017-12-20.12:26:34 |
|
Someone want to set up an experiment for this? The code change to test this is
simple, only a few lines. Michael or I can give pointers to the necessary
changes if it helps.
This is only relevant for configurations involving lazy search and preferred
operators, so I suggest we test lazy search with the FF heuristic and preferred
operators, lama-first, and lama (separately).
|
msg6756 (view) |
Author: silvan |
Date: 2017-12-20.12:23:08 |
|
Moved summary provided by Michael to Change Note.
|
msg6755 (view) |
Author: silvan |
Date: 2017-12-20.12:22:51 |
|
An introduction of EvaluationContext in issue77 has caused a change in the
behavior of
the lazy search.
Before: reward_progress() would be invoked every time there was a new best
heuristic
value found, including the initial state.
After: reward_progress() is not invoked for the initial state, and otherwise the
same.
It is not clear which behavior is better, but the change does seem to be
unintended and
is due to a separate handling of the progress check for the initial state in the
newer
version.
|
|
Date |
User |
Action |
Args |
2018-07-30 14:30:57 | manuel | set | status: reviewing -> resolved messages:
+ msg7337 |
2018-07-24 12:40:04 | silvan | set | messages:
+ msg7314 |
2018-07-24 12:13:30 | jendrik | set | messages:
+ msg7313 |
2018-07-24 12:11:27 | malte | set | messages:
+ msg7312 |
2018-07-24 12:04:24 | mkatz | set | messages:
+ msg7311 |
2018-07-24 11:47:39 | manuel | set | messages:
+ msg7310 |
2018-07-12 14:17:42 | silvan | set | messages:
+ msg7301 |
2018-06-07 19:21:32 | manuel | set | messages:
+ msg7240 |
2018-06-07 12:49:34 | malte | set | messages:
+ msg7235 |
2018-06-07 12:47:28 | malte | set | messages:
+ msg7234 |
2018-06-07 10:09:23 | silvan | set | status: in-progress -> reviewing messages:
+ msg7228 |
2018-06-06 11:57:10 | mkatz | set | messages:
+ msg7219 |
2018-06-06 10:53:24 | manuel | set | messages:
+ msg7218 |
2018-06-06 10:52:54 | silvan | set | messages:
+ msg7217 |
2018-06-06 10:44:43 | silvan | set | messages:
+ msg7216 |
2018-06-06 09:26:16 | manuel | set | messages:
+ msg7215 |
2018-06-04 12:22:42 | manuel | set | messages:
+ msg7155 |
2018-06-04 12:15:20 | silvan | set | messages:
+ msg7154 |
2018-06-04 12:13:52 | manuel | set | messages:
+ msg7153 |
2018-06-04 12:09:06 | silvan | set | messages:
+ msg7152 |
2018-06-04 12:06:35 | manuel | set | messages:
+ msg7151 |
2018-06-04 10:38:34 | manuel | set | status: chatting -> in-progress nosy:
+ manuel assignedto: manuel |
2017-12-20 12:27:40 | mkatz | set | messages:
+ msg6759 |
2017-12-20 12:27:32 | malte | set | messages:
+ msg6758 |
2017-12-20 12:26:34 | malte | set | messages:
+ msg6757 |
2017-12-20 12:23:39 | jendrik | set | nosy:
+ jendrik |
2017-12-20 12:23:08 | silvan | set | messages:
+ msg6756 |
2017-12-20 12:22:51 | silvan | set | status: unread -> chatting nosy:
+ malte, mkatz, silvan messages:
+ msg6755 summary: An introduction of EvaluationContext in issue77 has caused a change in the behavior of
the lazy search.
Before: reward_progress() would be invoked every time there was a new best heuristic
value found, including the initial state.
After: reward_progress() is not invoked for the initial state, and otherwise the same.
It is not clear which behavior is better, but the change does seem to be unintended and
is due to a separate handling of the progress check for the initial state in the newer
version. -> |
2017-12-20 12:21:57 | mkatz | create | |