Conversation
timothy-glover
requested review from
hpritchett-dstl and
nperree-dstl
and removed request for
a team
|
Docs appear to be failing on an unrelated example |
nperree-dstl
left a comment
nperree-dstl
left a comment
There was a problem hiding this comment.
I think this is a really interesting reward function to add! It does take a bit of effort to get your head around though as it's quite complicated. I've left some comments on where the documentation could be improved but longer term an example showcasing it's use would be great.
The only thing is I couldn't work out why the docs are failing, not sure if it's related to what you've done or not...
| points in the environment which each have an information value (information state). | ||
| The information state is decayed through time, with :attr:`decay_rate`, |
There was a problem hiding this comment.
I think I found the term "information state" confusing because of the existing InformationState object in Stone Soup. It might be worth clarifying the difference somewhere.
| doc="Reference level of information to achieve in each grid cell. " | ||
| "Vector of shape (1, n) where n is the number of cells. This should " | ||
| "match :attr:`environment_cells`." |
There was a problem hiding this comment.
This description could be improved perhaps - from our conversation this is how much information you are seeking to achieve in each grid and can be used to define priority areas? I'm not sure I understood that at first reading.
| ) | ||
|
|
||
| sensing_info_func: Callable = Property( | ||
| doc="Function describing the information from a sensing action." |
There was a problem hiding this comment.
Information gained? What are the required inputs/outputs of this function?
2 participants
This pull request introduces the
InformationCoverageRewardclass which implements an information coverage control objective for the sensor management module. The reward works with an information state which spatially quantifies the level of information at discrete points throughout the environment. The class rewards minimisation of the difference between the information state and a user defined reference information state multiplied into, if triggered, a target density or a uniform density, depending on whether a target is tracked or whether it is being searched for. Information here is a quantification of the area covered by sensors, and is left to the user to define for generality across sensor types and applications.The class requires that the information state is stored in track metadata in this implementation as the state should be carried through time as it evolves according to the information model.
The reward has also only been tested with a single target scenario but the concept is applicable to multitarget tracking problems and the implementation should allow this.
The new reward class has been incorporated into existing sensor manager tests. There is a problem with it being used with the
OptimizeBasinHoppingSensorManagerso assertions for this sensor manager andInformationCoverageRewardare skipped until the problem is resolved.