/// <summary>
/// Returns a flat collection of states/actions and their transition states.
/// </summary>
/// <param name="states">State matrix.</param>
/// <param name="actions">Action label vector.</param>
/// <param name="statesP">Transition states matrix.</param>
/// <param name="properties">(Optional) Feature summary.</param>
/// <param name="discretizer">Disretization function to apply for reducing states.</param>
/// <returns></returns>
public static Tuple <IEnumerable <IState>, IEnumerable <IAction>, IEnumerable <IState> > GetStates(Matrix states,
Vector actions, Matrix statesP, Math.Summary properties, IDiscretizer discretizer)
{
Math.Summary summary = (properties != null ? properties : Math.Summary.Summarize(states));
var slist = new IState[states.Rows];
var alist = new IAction[actions.Length];
var splist = new IState[statesP.Rows];
for (int i = 0; i < states.Rows; i++)
{
slist[i] = MDPConverter.GetState(states[i], summary, discretizer);
splist[i] = MDPConverter.GetState(statesP[i], summary, discretizer);
alist[i] = MDPConverter.GetAction(actions[i], slist[i].Id, splist[i].Id);
}
return(new Tuple <IEnumerable <IState>, IEnumerable <IAction>, IEnumerable <IState> >(slist, alist, splist));
}
/// <summary>
/// Returns a graph of MDP States from the States matrices and Action label vector.
/// </summary>
/// <param name="states">State matrix.</param>
/// <param name="actions">Action label vector.</param>
/// <param name="statesP">Transition states matrix.</param>
/// <param name="reward">Reward value vector.</param>
/// <param name="properties">Feature properties from the original set.</param>
/// <param name="discretizer">Discretization function for generating unique state identifiers.</param>
/// <returns>IEnumerable<IMDPState></returns>
public static IEnumerable <MDPState> GetStates(Matrix states, Vector actions, Matrix statesP, Vector reward,
Math.Summary properties, IDiscretizer discretizer)
{
Math.Summary summary = (properties != null ? properties : Math.Summary.Summarize(states));
discretizer.Initialize(states, summary);
var sdist = new Dictionary <double, MDPState>();
var adist = new Dictionary <string, double>();
var results = new Dictionary <double, MDPState>();
for (int i = 0; i < states.Rows; i++)
{
double sid = discretizer.Discretize(states[i], summary);
if (!sdist.ContainsKey(sid))
{
sdist.Add(sid, MDPConverter.GetState(states[i], summary, discretizer));
results.Add(sid, sdist[sid]);
}
double tsid = discretizer.Discretize(statesP[i], summary);
MDPState tstate = (sdist.ContainsKey(tsid) ? sdist[tsid] : MDPConverter.GetState(statesP[i], summary, discretizer));
if (!sdist.ContainsKey(tsid))
{
sdist.Add(tsid, tstate);
}
string key = GetActionKey((int)sid, (int)tsid);
if (!adist.ContainsKey(key))
{
adist.Add(key, 1);
}
else
{
adist[key]++;
}
sdist[sid].Successors.Add(new MDPSuccessorState(MDPConverter.GetAction(actions[i], (int)sid, (int)tsid), 0, tstate, reward[i]));
if (results.ContainsKey(tsid))
{
results.Remove(tsid);
}
}
foreach (var state in sdist.Values)
{
double sum = state.Successors.Sum(s => adist[GetActionKey(state.Id, s.State.Id)]);
foreach (var successor in state.Successors)
{
var key = GetActionKey(state.Id, successor.State.Id);
((Action)successor.Action).Probability = adist[key] / sum;
}
}
// return starting states
return(results.Values);
}