Simulation Standard

Important note: In most cases, users need not implement their own simulators. Several simulators that are compatible with the standard in this document are implemented in the POMDPSimulators package and allow interaction from a variety of perspectives. Moreover CommonRLInterface.jl provides an OpenAI Gym style environment interface to interact with environments that is more flexible in some cases.

In order to maintain consistency across the POMDPs.jl ecosystem, this page defines a standard for how simulations should be conducted. All simulators should be consistent with this page, and, if solvers are attempting to find an optimal POMDP policy, they should optimize the expected value of r_total below. In particular, this page should be consulted when questions about how less-obvious concepts like terminal states are handled.

POMDP Simulation

Inputs

In general, POMDP simulations take up to 5 inputs (see also the simulate docstring):

  • pomdp::POMDP: pomdp model object (see POMDPs and MDPs)
  • policy::Policy: policy (see Solvers and Policies)
  • up::Updater: belief updater (see Beliefs and Updaters)
  • b0: initial belief (this may be updater-specific, such as an observation if the updater just returns the previous observation)
  • s: initial state

The last three of these inputs are optional. If they are not explicitly provided, they should be inferred using the following POMDPs.jl functions:

Simulation Loop

The main simulation loop is shown below. Note that the isterminal check prevents any actions from being taken and reward from being collected from a terminal state.

Before the loop begins, initialize_belief is called to create the belief based on the initial state distribution - this is especially important when the belief is solver specific, such as the finite-state-machine used by MCVI.

b = initialize_belief(up, b0)

r_total = 0.0
d = 1.0
while !isterminal(pomdp, s)
    a = action(policy, b)
    s, o, r = @gen(:sp,:o,:r)(pomdp, s, a)
    r_total += d*r
    d *= discount(pomdp)
    b = update(up, b, a, o)
end

In terms of the explicit interface, the @gen macro above expands to the equivalent of:

    sp = rand(transition(pomdp, s, a))
    o = rand(observation(pomdp, s, a, sp))
    r = reward(pomdp, s, a, sp, o)
    s = sp

MDP Simulation

Inputs

In general, MDP simulations take up to 3 inputs (see also the simulate docstring):

The last of these inputs is optional. If the initial state is not explicitly provided, it should be generated using

Simulation Loop

The main simulation loop is shown below. Note again that the isterminal check prevents any actions from being taken and reward from being collected from a terminal state.

r_total = 0.0
disc = 1.0
while !isterminal(mdp, s)
    a = action(policy, s)
    s, r = @gen(:sp,:r)(mdp, s, a)
    r_total += d*r
    d *= discount(mdp)
end

In terms of the explicit interface, the @gen macro above expands to the equivalent of:

    sp = rand(transition(pomdp, s, a))
    r = reward(pomdp, s, a, sp)
    s = sp