This is the code for the intelligent prospector mineral exploration work conducted at Stanford University in a collaboration with SISL and SCERF. The source code provides a simulated mineral exploration problem and an implementation of the proposed sequential solution framework. The scripts needed to run the published experiments and additional trials are included in the scripts directory. All experiments were run in Julia 1.6.x. The specific versions of the required support packages are listed in the project.toml.
The package can be installed using the standard Julia package manager utility add function. To add the package from a local source, simply run the command (from within the package manager environment)
] add /PATH_TO_SOURCE
alternatively, you may navigate to the home directory of the source and run
] add .
If you would like to develop within the code base, we recommend building the source directly within a virtual environment. The requireed packages can be installed using the Julia build command from the source home directory.
All of the source code is in the src directory. The MineralExploration.jl file defines the main module strucutre. In this file, the source for all of the exposed structures and functions can be located. The parameters directory contains the data used to fit the variograms in the published experiments. The scripts directory contains the scripts to run the expeirments and tests.
The solve_pomcpow.jl script provides the best example of how to use this codebase. Key snippets are provided here.
using POMDPs
...
using MineralExploration
# 1) Build problem model
mainbody = SingleFixedNode()
m = MineralExplorationPOMDP(max_bores=MAX_BORES, delta=GRID_SPACING+1, grid_spacing=GRID_SPACING,
mainbody_gen=mainbody, max_movement=MAX_MOVEMENT)
# , geodist_type=GSLIBDistribution)
initialize_data!(m, N_INITIAL)
ds0 = POMDPs.initialstate_distribution(m)
# 2) Sample initial state
s0 = rand(ds0)
# 3) Create belief
up = MEBeliefUpdater(m, 1000, 2.0)
println("Initializing belief...")
b0 = POMDPs.initialize_belief(up, ds0)
println("Belief Initialized!")
ore_maps = [p.ore_map for p in b0.particles];
# 4) Create POMCPOW instance
next_action = NextActionSampler()
solver = POMCPOWSolver(tree_queries=10000,
check_repeat_obs=true,
check_repeat_act=true,
next_action=next_action,
k_action=2.0,
alpha_action=0.25,
k_observation=2.0,
alpha_observation=0.1,
criterion=POMCPOW.MaxUCB(100.0),
final_criterion=POMCPOW.MaxQ(),
# final_criterion=POMCPOW.MaxTries(),
estimate_value=0.0
# estimate_value=leaf_estimation
)
planner = POMDPs.solve(solver, m)
# 5) Run simulation
b_new = nothing
a_new = nothing
discounted_return = 0.0
B = [b0]
AE = Float64[]
ME = Float64[]
println("Entering Simulation...")
for (sp, a, r, bp, t) in stepthrough(m, planner, up, b0, s0, "sp,a,r,bp,t", max_steps=50)
...
@show t
@show a.type
@show a.coords
@show r
@show sp.stopped
@show bp.stopped
volumes = [sum(p.ore_map .>= m.massive_threshold) for p in bp.particles]
mean_volume = mean(volumes)
std_volume = std(volumes)
volume_lcb = mean_volume - 1.0*std_volume
push!(B, bp)
@show mean_volume
@show std_volume
@show volume_lcb
...
discounted_return += POMDPs.discount(m)^(t - 1)*r
end
In the above code, ... denotes code from the original script that was omitted for brevity here.
Note: To save data, the defalut location is a data folder in the source code home directory. To change this, change the file path strings referenced in the script.