D* pathfinding algorithm

Never heard of D*... you sure you're not looking for the A* algorithm?

Quote:

Original post by JBourrie Never heard of D*... you sure you're not looking for the A* algorithm?

D* is a modification of A* to take dynamic elements into account. I don't know of any good freely available articles about it that doesn't go into a lot of details. If you have some theoretical foundation you might want to see the original paper, but it's definitely too heavy for most people who don't know D* yet.

I stand corrected :)

Seems odd that a Google search doesn't even bring up a mention of the algorithm (or at least my search case didn't).

Although Anthony Stentz wrote the first D* paper, some of the most significant follow up research has been conducted my Sven Keonig (http://idm-lab.org/publications.html) and Maxim Likhachev (http://www.cs.cmu.edu/~maxim/). From what I can tell it is now considered more common not to use the Stentz implementation, but to use the LPA* derivatives which are easier to code and analyse.

D* Light papers:
http://idm-lab.org/bib/abstracts/papers/aaai02b.pdf, http://www.cs.cmu.edu/~maxim/docs/dlitemap_iros02.pdf

I would strongly recommend that you implement other search algorithms before starting to work with D* (probably A* and LPA* in that order), it's always nice to have a reference implementation of another search that you can compare the results to. D* can be quite hard to get right.

Depending on the applications on the applications (practical robotics?) you might also want to have a look at Jur van den Berg and Dave Ferguson's research on the Anytime D* (http://gs2045.sp.cs.cmu.edu/research.php).

I thank everyone for their replies.

I wrote my honours project on the future of pathfinding algorithms in games, and D* came up. It was difficult to implement, but it's major downfall was that in order for it to function properly, each agent needed it's own copy of the map. This is simply unacceptable for games, as there can be huge numbers of agents in a search space. the space complexity (or memory requirements) were orders of magnitude higher than that of A*, which is quite heavy on memory.

My conclusion was that A* was still the best choice, and there are plenty of tweaks and optimisations for A*.

BUT, I wont say that no-one could make it work.

If you do, let me know.

If you want a copy of my honour report (although, the best thing you'd get from it would be my references page), let me know.

I've just started reading about D* Lite and if possible I'll make an implementation of the algorithm plus A*. I implemented A* a long time ago and decided to implement it once again (code has a habit of disappearing) when i came across a mention of D*.

I'm very curious as to your statement that each agent needs it's own copy of the map in order for the algorithm to work, so I'd love it if you could send the report you did to the email address that I've PM:ed you.

Yes, if you're using D* for multiple agents you will need multiple copies of some parts of the graph.

D* is only relevant when you are updating the costs of the edges. This means that you have incoming data being delivered (perhaps via sensors or simulated sensors), whilst you are moving - an online solution. This means that unlike A*, you can't just create your plan and just recycle the graph - an offline solution.

For multiple agents planning simultaneously that means each must maintain their own graph with the appropriate data (g, rhs, etc.) on each vertex. The amount of memory required is proportional only to the number of vertices in the graph, not the edges. For a dense uniform grid this may be expensive, but for well designed undirected graphs it can be a done reasonably small amount of space.

Cybrosys, mind if I ask what the application is? I might have a better solution for you.

Cheers,

Timkin

There is no specific application, i was planning on implementing the algorithm as dynamically/abstract as possible for future usage.