Near-Optimal Search in Continuous Domains. Ieonig, Lambert, Shoham, Brafman. AAAI 2007

  1. Concerned with search in deterministic continuous state and action spaces, time is discrete
  2. Deals with goal-finding in the presence of a heuristic function, paths must be bounded on their length, d
  3. Getting exactly optimal results in continuous domains is only possible at the limit, so instead planning is done to epsilon-accuracy
  4. If the goal state cannot be found in d steps, it will return the first d steps of a path that may reach the goal optimally, as well as its lower bound, which can be done because of the heuristic
  5. Argue that use of forward search allows for more efficient algorithms because of pruning
  6. Discuss algorithms from the motion planning literature: Probabalistic Roadmaps, and Rapidly Exploring Random Trees
    1. Issues here are motion planning is only concerned with finding a path from start to goal, and does not care about optimality
    2. Additionally, motion planning algorithms assume continuous time, and when time is discrete the constraints are non-differentiable
  7. In general, the literature on continuous state, action, discrete time is very limited
  8. Transitions, cost, and heuristic functions must be Lipshitz
  9. Algorithm basically does Lipshitz optimization over the cost of paths starting from the range of actions at each node in the rollout tree
  10. Samples are done in a rollout-manner, as opposed to a depth-first blowout of the samples, so M, the action -> cost function for each node is updated for all nodes encountered during a rollout after it completes
  11. They discuss how to do the updates based on the Lipshitz constraints in higher dimensions than 1, where there resulting constraints are an envelope of hyper-cones, but they estimate this with rectangles, which is what I have seen in the literature on pure Lipshitz optimization.  This can be done in a way that is always pessimistic, so it doesn’t really mess up the results (wont prevent an epslion-optimal result to be found, but it may take more samples than if cones were actually used)
  12. Guarantees (with proofs):
    1. The algorithm terminates
    2. At termination, it returns a lower bound on the cost of any plan
    3. If the plan returned is complete, it is at most epsilon worse than optimal
    4. If the plan returned is partial, the beginning of the plan returned in the best case will be epsilon worse than optimal
  13. The paper is primarily theoretical, but gives fair space to how it may actually be implemented (for example, approximating cones with rectangles)
  14. Discuss importance of extending to to stochastic setting)

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: