1: List five ideas/techniques that you learned in this course
that you thought were most interesting?
|
Student |
Answer |
|
Dan
Bryce |
1)
Refinement
Planning Generalization 2)
Being able to
understand PO Planning, 3)
The discussion
on local Search Techniques, but would have liked a more general discussion,
not just LPG 4)
The discussion
of TCN’s. 5)
Planning under
uncertainty discussions. |
|
J.
Benton |
1)
EBL and memoization 2)
Using the
planning graph to derive heuristics. 3)
Metric Temporal
Planning 4)
Conformant
Planning 5)
Progression, Regression
and causal proofs. |
|
Mike
Rush |
1)
Explanation
Based Learning 2)
Graphplan 3)
Knowledge Based
Planning 4)
Heuristics 5)
Markov Decision
Problems |
|
Xin
Z. |
1)
Planning Graphs 2)
Planning
Heuristics 3)
Conditional
Planning 4)
Planning with
incompleteness 5)
Scheduling |
|
Preetha
A. |
1)
Partial Order
planning: not committing to position. 2)
Posing planning
as CSP and seeing that they actually speed up search 3)
The idea of
planning graphs for reachability as well as
heuristics 4)
The local
search idea and seeing how it worked to give faster search times but not
optimal plans 5)
Seeing that
instead of committing to one approach if we combine advantages of different
approaches in planning, it turns out that better ways to solve problems are
obtained. |
|
Corey
M. |
1)
Conditional
planning 2)
Conformant
planning 3)
MDPs 4)
Metric Temporal
Planning 5)
Temporal
Constrain Networks |
|
Monika
S |
1)
Least-commitment-
It provides a lot o freedom when it comes to deciding a position of an
action, especially in early stages of creating a plan. 2)
Memos in Graphplan search. It was great way of pruning search. 3)
Planning with
Knowledge-based planners. It can do wonders in improving plan quality. 4)
Capturing
n-sized mutexes and how that captures n-ary relations between state variables. 5)
I liked
creating my own domain in PDDL 2.1, even though the lack of support for adl was very limiting. |
|
Menkes
v. |
1)
The graphplan algorithm 2)
Heuristic
estimation in Graphplan 3)
POP Algorithm 4)
DDB and EBL 5)
Metric temporal
planning approaches by SAPA and LPG |
|
Josh
R. |
1)
Conformant
planning 2)
The value of
graph plan and similar ideas even in non classical domains 3)
How much easier
(well for the user anyway) metric temporal planning is compared to how hard
it sounds… this is a big hurray to PDDL 2.1 4)
The value of
using sets of states compared to individual states. 5)
The cost of
using sets of states compared to individual states. |
2: List five ideas/technqiues that you
thought got over-play and were not worth the time we spent on them.
|
Student |
Answer |
|
Dan
Bryce |
1)
Too much
testing of planners (BB, GPCSP, Sapa, LPG) in
projects. 2)
Proofs. Writing
causal proof for homework and midterm was too much. 3)
LPG. Nice to
learn about, but we didn’t go into such detail with other leading planning
approaches. 4)
Difference
between conformant and conditional planning. You spent too much time
outlining the differences in the problem definitions and not enough on the
solutions techniques. 5)
Not enough on
learning, just the EBL/DDB stuff. |
|
J.
Benton |
1)
Knowledge Based
Planning 2)
CSP encoding of
Graphplan (good to discuss but I think overplayed) 3)
Partial order
planning (spent too much time on) 4)
Compilation to
IP 5)
Would have
liked to have tarted temporal planning sooner so we
could cover it for a longer period of time |
|
Mike
Rush |
I didn’t think any topic
was over covered. I would have liked to spend more time on temporal problems
rather than classical ones. For me, they seemed to be more “real life” even
though the methods are not as developed. |
|
Xin
Z. |
1)
SAT encoding 2)
ILP encoding 3)
CSP encoding 4)
SAPA 5)
ZENO |
|
Preetha
A. |
1)
Initial parts
of progression and regression search, too much time
was spent on it. 2)
More on metric
temporal planning could have been done. 3)
Knowledge based
planners: What sorts of problems do they actually solve? |
|
Corey
M. |
1)
Basic
Progression 2)
Basic
Regression 3)
Partial Order
Planning 4)
Graphplan 5)
CSP/SAT |
|
Monika
S |
1)
Converting
planning graphs to CSP and especially SAT encoding. I thought that there was
too much busy-work involved with it. 2)
I think that
too much time was spent on discussion how to fix open conditions in different
planning schemas. 3)
Empirical
problems. Running sets of problems with different planners. I think it would
have been more productive if there was less busy-work and more per-example
explanations, i.e. this planner is generally better in these types of
problems, but it failed in problem 1, because … 4)
There was too
much busy work with making mutexes and generally
too much discussion about mutexes relations. I do
realize though, that it was important in capturing negative relations between
actions and in pruning search. 5)
There was to omuch time spen on serial plans
versus parallel plans, i.e. we spent a lot of time talking about the blocks
world. It’s great and easy to use in different examples, but I would like to
see more examples of domains with parallel plans. |
|
Menkes
v. |
1) Hmm, difficult.
Actually, I felt like we went through a lot of things very quickly,
especially in the first couple of weeks. I strongly feel that he time spent
on classical planning is time well spent. It gives you a solid basis to
tackle more advanced planning problems like MTP and planning under uncertainty. |
|
Josh
R. |
Most things seemed to
lead into something else eventually, even if they felt overplayed at the
time, looking back I don’t think it would be very fair to call any of them
over played. |
3: List any
ideas that you thought would be covered in the course, and would have liked
them to see covered, but finally were not.
|
Student |
Answer |
|
Dan
Bryce |
1)
It would have
been nice to talk more about applications (domains) that are used in
industry. Plus, where the demand is. 2)
I would have
liked to learn more about monitoring plan execution. (RAX) 3)
It would have
been nice to have a general overview of algorithms that are used, and in what
scenarios they are most applicable. 4)
I wanted to
hear more (or something) about multi-agent systems. 5)
I was
disappointed that we were not assigned individual semester projects, in place
of homogenous, smaller projects. |
|
J.
Benton |
1)
I was expecting
a little more information about learning techniques, especially the monte carlo methods of
reinforcement learning |
|
Mike
Rush |
1) I would have liked
some discussion of how all these methods relate to how humans solve these
planning problems (which we seem to do all the time). Even just a survey of
hypotheses with brief discussion would have been nice. |
|
Xin
Z. |
1)
The class is
called “Planning and Learning”, but planning takes 90% of the class time, and
learning only covered a little bit. 2)
I would like to
see more scheduling part. 3)
Also, we may
need to spend one class on CSP, since not everyone learn it before 4)
So, my
suggestion is (might be wrong, but just my opinion) that planning 50% and
rest other topics such as scheduling, learning. |
|
Preetha
A. |
1)
Fully
concentrated on planning, except EBL methods. There wasn’t anything on
learning. 2)
However the
time was too short and you packed as much as possible. I worked hard and
learned so much about planning from a level that I did not commit to a single
method. Encourages to think on different approaches. |
|
Corey
M. |
I would have liked to see
coverage of more advanced topics in Planning, such as disjunctive goals.
Also, it would have been interesting to have leed
of more applications of planning in the real world. |
|
Monika
S |
I would like to see more
planning based on agents acquiring knowledge from its experiences. I was also
hoping for planning in dynamic domains. It would have been nice to be able to
use planners that are able of handling conditional effects, negative
preconditions, disjunctive preconditions, etc. I also thought I would
see more elaborate planning used in strategy games and simulators, like the
“Massive” (crowd simulator) |
|
Menkes
v. |
When the notion of lifted
actions was introduced I was looking forward to see a full class (maybe two)
on this particular topic. It is one of a few topics I am still not very
comfortable with. |
|
Josh
R. |
The probabilistic stuff
could have been fun, but I feel we covered a lot of stuff for one semester as
is so that’s ok. |
4: List five "non-obvious" (to you) details or
cross-connections that _you_ were able to appreciate over the semester
|
Student |
Answer |
|
Dan Bryce |
1) The connections
between Graphplan, GPCSP, and Blackbox
were good to learn. 2) I liked the three
categories for planning under uncertainty: problems, solutions, and success
measures. 3) It was nice to see the
101 ways to that the planning graph can be used in different planning
scenarios. 4) I liked the
explanation of the continuum between planning and scheduling. 5) The connections
between SAT, CSP, and ILP were interesting, but I would have liked a deeper
examination. |
|
J.
Benton |
1)
Calculating the
slack between two events can be used as a heuristic for finding an ordering
of events, using de Floyd-Warshall algorithm to
tighten constraints on simple temporal constraint network. 2)
Being unable to
observe your state and having non-deterministic actions are
the “same thing”. 3)
Serial planning
graph finds the shortest plan in terms of numbers of action but normal
parallel graph finds the shortest plan in terms of number of levels. 4)
Solving a
compiled to CSP grpahplan allows the ability to
work on different levels of the graphplan at the
same time. 5)
It is possible
that graphplan will not find the optimal parallel
plan to a problem. |
|
Mike
Rush |
1)
Planning is
search, search and search. 2)
Search spaces
come in different varieties (state, belief, plans, etc). 3)
MDP applicability
to finances 4)
The limits on
classical planning and how those limits affect the range of problems which
can be solved. |
|
Xin
Z. |
1)
I learned
planning heuristics. 2)
I also learned
different kinds of techniques about planning. At beginning of the class, I
have little idea about planning, but now I think I know it, even though not
completely, but at least got some ideas. 3)
Temporal
planning is very interesting. 4)
Planning with
incompleteness is hard, it seemed that there is not much work have been done
compared with other topics we covered. But the introduction of this idea is
good. I want to hear more about it, very good presentation of this part. 5)
Conditional
planning is also god knowledge that I learned from this class. It’s
interesting and Dr. Rao gave good class in this part. |
|
Preetha
A. |
1)
Mutexes doing binary mutexes till level off does not guarantee a plan of that
length (because we do not calculate 3,4 mutexes,
etc) 2)
Regression
search looks faster and more efficient up front, but actually we search in
sets of states many of which may not even be legal, so it is costly. 3)
What encoding
to follow CSP/SAT/IP depends of the problem too, there is not clear winner. 4)
Notion of
preconditions and effects of actions that wasn’t quite true while moving into
conformant/conditional planning in belief states. It becomes more like do the
action and if the precondition is true the effect will happen, if not nothing
happens. |
|
Corey
M. |
1)
Reachability heuristics 2)
EBL 3)
Bellman
equations and their convergence 4)
Search in
AND/OR Graphs 5)
Non-classical
planning in general. |
|
Monika
S |
1)
Earlier
planners could have solved the Sussman-anomaly
problem if the goal state description would have been complete. Intuitively
would have said that the more relaxed the goal state, the easier would be to
satisfy it. 2)
Memos are n-ary mutexes. For some reason it
was not obvious to me, but it was really helpful in realizing a link between
different searches. 3)
If a set of
subgoals have mutual exclusion relations among them and appear together for
the first time at step n without any pair being mutex,
does not guarantee that thre is a plan to achieve
these subgoals in an n-step plan, because only mutexes
between pairs are being considered. There might be 3,4,5,
..-ary mutexes. The cross
connection is that this can be compared to arc consistency does not guarantee
there is solution, n-consistency (for n step plan) does. 4)
A partial order
planner is a least commitment regression planner. 5)
Given that
planners had very hard time finding solutions when plan required temporarily
getting away from the goal, one of the hardest problems to solve for the
planners would be the rubik’s cube. The evil part
is that (quoting from http://www.rubiks.com/):
“Theoretically the shortest possible path to solving Rubik’s cube from any
scrambled position is as few as 22 twists . So far
no one has succeded in demonstrating this method. |
|
Menkes
v. |
1)
The planning
circle of Graphplan-CSP-SAT-IP 2)
EBL and
memorization 3)
Solving
classical planning as MDP |
|
Josh
R. |
1)
It’s not really
that non-obvious , but I had not realized the connection between no-goods and
mutex, until you said it and then I said “duh” 2)
The actual
speed and quality of a greedy planner (LPG) really caught m off guard. |