Implementation of different algorithms and policies for solving MDPs.
MDPs are specified directly in C++ using the supplied template libraries.

Algorithms, policies and heuristics are specified as string requests
passed to the dispatcher.  See examples, like sailing, on how to implement
this.

Typical requests:

Algorithms for solving the MDP form the initial state
(all parameters get default values, specify if you want to change):

algorithm=value-iteration(epsilon=<float>,max-number-iterations=<integer>,heuristic=<request>,seed=<integer>)
algorithm=improved-lao(epsilon=<float>,heuristic=<request>,seed=<integer)
algorithm=hdp(epsilon=<float>,heuristic=<request>,seed=<integer)
algorithm=ldfs(epsilon=<float>,heuristic=<request>,seed=<integer)
algorithm=ldfs-plus(epsilon=<float>,heuristic=<request>,seed=<integer)
algorithm=plain-check(epsilon=<float>,heuristic=<request>,seed=<integer)
algorithm=lrtdp(epsilon=<float>,heuristic=<request>,epsilon_greedy=<float>,seed=<integer)
algorithm=standard-lrtdp(epsilon=<float>,heuristic=<request>,epsilon_greedy=<float>,seed=<integer)
algorithm=uniform-lrtdp(epsilon=<float>,heuristic=<request>,epsilon_greedy=<float>,seed=<integer)
algorithm=bounded-lrtdp(epsilon=<float>,heuristic=<request>,bound=<integer>,epsilon_greedy=<float>,seed=<integer)
algorithm=simple-a*(heuristic=<request>,seed=<integer)

// lrtdp is alias for standard-lrtdp
// for bounded-lrtdp(), it is not a good idea to use default value for bound
// simple-a* is only good for deterministic MDPs (i.e. OR graphs)
// Don't assign values to bound and epsilon-greedy unless you know what you are doing.

// Default values:

seed = 0
epsilon = 0.0 // not a good idea to use this default value
heuristic = null // translates into zero estimates
max-number-iterations = max
bound = max
epsilon-greedy = 0.0


// Heuristics (used in algorithms and policies):

zero()
min-min(algorithm=<request>)
optimal(algorithm=<request>)
scaled(heuristic=<request>,weight=<float>))

// Parameters for heuristics are mandatory


// Policies for online-planning from initial state:

policy=random()
policy=greedy(optimistic=<boolean>,random-ties=<boolean>,caching=<boolean>,heuristic=<request>)
policy=optimal(algorithm=<request>)
policy=rollout(width=<integer>,depth=<integer>,nesting=<integer>,policy=<request>)
policy=uct(width=<integer>,horizon=<integer>,parameter=<float>,random-ties=<boolean>,policy=<request>)
policy=aot(width=<integer>,horizon=<integer>,probability=<float>,expansions-per-iteration=<integer>,random-ties=<boolean>,policy=<request>,heuristic=<request>)
policy=finite-horizon-lrtdp(horizon=<integer>,max-trials=<integer>,labeling=<boolean>,random-ties=<boolean>,heuristic=<request>)