sample_belief_space function

Sample from the Belief Space

Sample points from belief space using a several sampling strategies.

sample_belief_space(model, projection = NULL, n = 1000, method = "random", ...)

Arguments

• model: a unsolved or solved POMDP .
• projection: Sample in a projected belief space. See projection() for details.
• n: size of the sample. For trajectories, it is the number of trajectories.
• method: character string specifying the sampling strategy. Available are "random", "regular", and "trajectories".
• ...: for the trajectory method, further arguments are passed on to simulate_POMDP(). Further arguments are ignored for the other methods.

Returns

Returns a matrix. Each row is a sample from the belief space.

Details

The purpose of sampling from the belief space is to provide good coverage or to sample belief points that are more likely to be encountered (see trajectory method). The following sampling methods are available:

• 'random' samples uniformly sample from the projected belief space using the method described by Luc Devroye (1986). Sampling is be done in parallel after a foreach backend is registered.
• 'regular' samples points using a regularly spaced grid. This method is only available for projections on 2 or 3 states.
• "trajectories" returns the belief states encountered in n trajectories of length horizon starting at the model's initial belief. Thus it returns n x horizon belief states and will contain duplicates. Projection is not supported for trajectories. Additional arguments can include the simulation horizon and the start belief which are passed on to simulate_POMDP().

Examples

data("Tiger")

# random sampling can be done in parallel after registering a backend.
# doparallel::registerDoParallel()

sample_belief_space(Tiger, n = 5)
sample_belief_space(Tiger, n = 5, method = "regular")
sample_belief_space(Tiger, n = 1, horizon = 5, method = "trajectories")

# sample, determine the optimal action and calculate the expected reward for a solved POMDP
# Note: check.names = FALSE is used to preserve the `-` for the state names in the dataframe.
sol <- solve_POMDP(Tiger)
samp <- sample_belief_space(sol, n = 5, method = "regular")
data.frame(samp, action = optimal_action(sol,  belief = samp), 
  reward = reward(sol, belief = samp), check.names = FALSE)
  
# sample from a 3 state problem
data(Three_doors)
Three_doors

sample_belief_space(Three_doors, n = 5)
sample_belief_space(Three_doors, n = 5, projection = c(`tiger-left` = .1))

if ("Ternary" %in% installed.packages()) {
sample_belief_space(Three_doors, n = 9, method = "regular")
sample_belief_space(Three_doors, n = 9, method = "regular", projection = c(`tiger-left` = .1))
}

sample_belief_space(Three_doors, n = 1, horizon = 5, method = "trajectories")

References

Luc Devroye, Non-Uniform Random Variate Generation, Springer Verlag, 1986.

See Also

Other POMDP: MDP2POMDP, POMDP(), accessors, actions(), add_policy(), plot_belief_space(), projection(), reachable_and_absorbing, regret(), simulate_POMDP(), solve_POMDP(), solve_SARSOP(), transition_graph(), update_belief(), value_function(), write_POMDP()

Author(s)

Michael Hahsler