pomdp (version 1.2.4)

add_policy: Add a Policy to a POMDP Problem Description

Description

Add a policy to a POMDP problem description allows the user to test policies on modified problem descriptions or to test manually created policies.

Usage

add_policy(model, policy)

Value

The model description with the added policy.

Arguments

model

a POMDP or MDP model description.

policy

a policy data.frame.

Author

Michael Hahsler

See Also

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

Other MDP: MDP(), MDP2POMDP, MDP_policy_functions, accessors, actions(), gridworld, reachable_and_absorbing, regret(), simulate_MDP(), solve_MDP(), transition_graph(), value_function()

Examples

data(Tiger)

sol <- solve_POMDP(Tiger)
sol

# Example 1: Use the solution policy on a changed POMDP problem
#            where listening is perfect and simulate the expected reward

perfect_Tiger <- Tiger
perfect_Tiger$observation_prob <- list(
  listen = diag(1, length(perfect_Tiger$states), 
    length(perfect_Tiger$observations)), 
  `open-left` = "uniform",
  `open-right` = "uniform"
)

sol_perfect <- add_policy(perfect_Tiger, sol)
sol_perfect

simulate_POMDP(sol_perfect, n = 1000)$avg_reward

# Example 2: Handcraft a policy and apply it to the Tiger problem

# original policy
policy(sol)
plot_value_function(sol)
plot_belief_space(sol)

# create a policy manually where the agent opens a door at a believe of 
#  roughly 2/3 (note the alpha vectors do not represent 
#  a valid value function)
p <- list(
data.frame(
  `tiger-left` = c(1, 0, -2),
  `tiger-right` = c(-2, 0, 1), 
  action = c("open-right", "listen", "open-left"),
  check.names = FALSE
))
p

custom_sol <- add_policy(Tiger, p)
custom_sol

policy(custom_sol)
plot_value_function(custom_sol)
plot_belief_space(custom_sol)

simulate_POMDP(custom_sol, n = 1000)$avg_reward