public static GreedyPolicy <TState, TAction> Create(
IProblem <TState, TAction> problem,
ValueTable <TState, TAction> valueTable,
IRewarder <TState, TAction> rewarder)
{
var greedyPolicy = new GreedyPolicy <TState, TAction>(problem);
foreach (var state in problem.AllStates())
{
var bestAction = FindBestAction(problem, state, valueTable, rewarder);
greedyPolicy._actions[state] = bestAction;
}
return(greedyPolicy);
}
FindOptimalPolicy <TState, TAction>(
IProblem <TState, TAction> problem,
IRewarder <TState, TAction> rewarder,
int evaluationSweepsPerPolicyUpdate = 1)
where TState : struct
where TAction : struct
{
const int maxIterations = 100;
var values = new ValueTable <TState, TAction>(problem);
IPolicy <TState, TAction> initialPolicy = new UniformRandomPolicy <TState, TAction>(problem);
values.Evaluate(initialPolicy, rewarder, evaluationSweepsPerPolicyUpdate);
var greedyPolicy = GreedyPolicy <TState, TAction> .Create(problem, values, rewarder);
for (var i = 0; i < maxIterations; i++)
{
values.Evaluate(greedyPolicy, rewarder, evaluationSweepsPerPolicyUpdate);
var newGreedyPolicy = GreedyPolicy <TState, TAction> .Create(problem, values, rewarder);
if (newGreedyPolicy.HasSameActionsAs(greedyPolicy))
{
Console.WriteLine($"Found optimal policy at iteration {i}");
break;
}
greedyPolicy = newGreedyPolicy;
if (i == maxIterations - 1)
{
Console.WriteLine($"Policy iteration did not converge by iteration {i}");
}
}
return(greedyPolicy, values);
}