The Partially Observable Monte Carlo Planning (POMCP) online solver for POMDPs.jl.
Silver, D., & Veness, J. (2010). Monte-Carlo Planning in Large POMDPs. In Advances in neural information processing systems (pp. 2164–2172). Retrieved from http://discovery.ucl.ac.uk/1347369/
This package replaces the deprecated POMCP.jl package. It has fewer features, but is much simpler.
In Julia 1.0 and higher, BasicPOMCP is available on the JuliaPOMDP registry (and this registry is required to automatically get the MCTS dependency):
using Pkg Pkg.add("POMDPs") using POMDPs POMDPs.add_registry() Pkg.add("BasicPOMCP")
using POMDPs, POMDPModels, POMDPSimulators, BasicPOMCP pomdp = TigerPOMDP() solver = POMCPSolver() planner = solve(solver, pomdp) for (s, a, o) in stepthrough(pomdp, planner, "sao", max_steps=10) println("State was $s,") println("action $a was taken,") println("and observation $o was received.\n") end
A minimal example showing how to implement a problem for the solver to use is here: https://github.com/JuliaPOMDP/BasicPOMCP.jl/blob/master/notebooks/Minimal_Example.ipynb
Solver options are set with keyword arguments to the
BasicPOMCP constructor. The options are described in the docstring which can be accessed with
The output of
?POMCPSolver is printed below, but may not be up to date.
Partially Observable Monte Carlo Planning Solver.
max_depth::IntRollouts and tree expension will stop when this depth is reached. default:
c::Float64UCB exploration constant - specifies how much the solver should explore. default:
tree_queries::IntNumber of iterations during each action() call. default:
estimate_value::AnyFunction, object, or number used to estimate the value at the leaf nodes. default:
- If this is a function
f(pomdp, s, h::BeliefNode, steps)will be called to estimate the value.
- If this is an object
estimate_value(o, pomdp, s, h::BeliefNode, steps)will be called.
- If this is a number, the value will be set to that number Note: In many cases, the simplest way to estimate the value is to do a rollout on the fully observable MDP with a policy that is a function of the state. To do this, use
default_action::AnyFunction, action, or Policy used to determine the action if POMCP fails with exception
- If this is a Function
f(pomdp, belief, ex)will be called.
- If this is a Policy
action(p, belief)will be called.
- If it is an object
default_action(a, pomdp, belief, ex)will be called, and if this method is not implemented,
awill be returned directly.
rng::AbstractRNGRandom number generator. default:
This solver does not reuse decision-making simulations for the belief update as in the original Silver and Veness paper. We have found that unweighted particle filtering approach to be unuseful in practice because
Instead, a custom belief updater, or an updater from ParticleFilters.jl should be used.
The search tree can be visualized with D3Trees.jl after running the
action_info() as in the example below. Note: tree_in_info must be set to true either as a solver option or as a keyword argument to action_info() for this to work (it is disabled by default because it can use a lot of memory).
using POMDPs using BasicPOMCP using POMDPModels using POMDPModelTools using D3Trees using Random pomdp = BabyPOMDP() solver = POMCPSolver(tree_queries=1000, c=10.0, rng=MersenneTwister(1)) planner = solve(solver, pomdp) a, info = action_info(planner, initialstate_distribution(pomdp), tree_in_info=true) inchrome(D3Tree(info[:tree], init_expand=3))
This should produce the image at the top of the README.
about 1 month ago