An implementation of the AR-DESPOT (Anytime Regularized DEterminized Sparse Partially Observable Tree) online POMDP Solver.
Tried to match the pseudocode from this paper: http://bigbird.comp.nus.edu.sg/m2ap/wordpress/wp-content/uploads/2017/08/jair14.pdf as closely as possible. Look there for definitions of all symbols.
Problems use the POMDPs.jl generative interface.
If you are trying to use this package and require more documentation, please file an issue!
On Julia v1.0 or later, ARDESPOT is in the Julia General registry
Pkg.add("POMDPs")
Pkg.add("ARDESPOT")
using POMDPs, POMDPModels, ARDESPOT
using POMDPTools
pomdp = TigerPOMDP()
solver = DESPOTSolver(bounds=(-20.0, 0.0))
planner = solve(solver, pomdp)
for (s, a, o) in stepthrough(pomdp, planner, "s,a,o", max_steps=10)
println("State was $s,")
println("action $a was taken,")
println("and observation $o was received.\n")
end
For minimal examples of problem implementations, see this notebook and the POMDPs.jl generative docs.
Solver options can be found in the DESPOTSolver
docstring and accessed using Julia's built in documentation system (or directly in the Solver source code). Each option has its own docstring and can be set with a keyword argument in the DESPOTSolver
constructor. The definitions of the parameters match as closely as possible to the corresponding definition in the pseudocode of this paper.
In most cases, the recommended way to specify bounds is with an IndependentBounds
object, i.e.
DESPOTSolver(bounds=IndependentBounds(lower, upper))
where lower
and upper
are either a number or a function (see below).
Often, the lower bound is calculated with a default policy, this can be accomplished using a DefaultPolicyLB
with any Solver
or Policy
.
If lower
or upper
is a function, it should handle two arguments. The first is the POMDP
object and the second is the ScenarioBelief
. To access the state particles in a ScenairoBelief
b
, use particles(b)
(or collect(particles(b))
to get a vector).
In most cases, the check_terminal
and consistency_fix_thresh
keyword arguments of IndependentBounds
should be used to add robustness (see the IndependentBounds
docstring for more info).
For the BabyPOMDP
from POMDPModels
, bounds setup might look like this:
using POMDPModels
using POMDPTools
always_feed = FunctionPolicy(b->true)
lower = DefaultPolicyLB(always_feed)
function upper(pomdp::BabyPOMDP, b::ScenarioBelief)
if all(s==true for s in particles(b)) # all particles are hungry
return pomdp.r_hungry # the baby is hungry this time, but then becomes full magically and stays that way forever
else
return 0.0 # the baby magically stays full forever
end
end
solver = DESPOTSolver(bounds=IndependentBounds(lower, upper))
Bounds need not be calculated independently; a single function that takes in the POMDP
and ScenarioBelief
and returns a tuple containing the lower and upper bounds can be passed to the bounds
argument.
D3Trees.jl can be used to visualize the search tree, for example
using POMDPs, POMDPModels, D3Trees, ARDESPOT
using POMDPTools
pomdp = TigerPOMDP()
solver = DESPOTSolver(bounds=(-20.0, 0.0), tree_in_info=true)
planner = solve(solver, pomdp)
b0 = initialstate_distribution(pomdp)
a, info = action_info(planner, b0)
inchrome(D3Tree(info[:tree], init_expand=5))
will create an interactive tree that looks like this:
DESPOT.jl was designed to exactly emulate the C++ code released by the original DESPOT developers. This implementation was designed to be as close to the pseudocode from the journal paper as possible for the sake of readability. ARDESPOT has a few more features (for example DESPOT.jl does not implement regularization and pruning), and has more compatibility with a wider range of POMDPs.jl problems because it does not emulate the C++ code.