// Initialise la map + l'agent
private void Initialisation()
{
this.agentGridWorld = new AgentGridWorld();
// Initialisation des actions possible
var rightAction = new MoveAction(new Vector2Int(1, 0), "DR");
var leftAction = new MoveAction(new Vector2Int(-1, 0), "DL");
var upAction = new MoveAction(new Vector2Int(0, 1), "DU");
var downAction = new MoveAction(new Vector2Int(0, -1), "DD");
// On les ajoutes dans une liste afin de definir tout les actions possible pour un etat donne
var actions = new List <Action>();
actions.Add(rightAction);
actions.Add(leftAction);
actions.Add(upAction);
actions.Add(downAction);
// On declare un dictionnaire afin d'y stocker les action possible pour un etat donne
Dictionary <Vector2Int, List <Action> > actionsDic = new Dictionary <Vector2Int, List <Action> >();
// Initialisation de la grid
grid = new GridCell[gridParameter.gridSize.x, gridParameter.gridSize.y];
for (int i = 0; i < gridParameter.gridSize.x; i++)
{
for (int j = 0; j < gridParameter.gridSize.y; j++)
{
// Initialisation du state par defaut et de la position de la grid
grid[i, j] = new GridCell()
{
position = new Vector2Int(i, j),
state = GridCell.GridState.Walkable
};
// Instanciation du visuel de la case
grid[i, j].visual = Instantiate(gridParameter.gridCellPrefab, new Vector3(i, j, 0.0f),
Quaternion.identity);
grid[i, j].visual.transform.SetParent(gridParent);
// Definition du reward initial
grid[i, j].r = -1.0f;
grid[i, j].v = 0.0f;
// Initialisation vide des actions pour le state [i, j]
var availableActions = new List <Action>();
actionsDic.Add(grid[i, j].position, availableActions);
}
}
// Definition du start state (pas utile pour le training de l'agent)
gridParameter.startState = grid[Random.Range(0, gridParameter.gridSize.x), Random.Range(0, gridParameter.gridSize.y)].position;
grid[gridParameter.startState.x, gridParameter.startState.y].state = GridCell.GridState.Start;
// Definition du end state (décommenter la ligne 132 et commenter la 133 pour que ce soit aleatoire)
// gridParameter.targetState = grid[Random.Range(0, gridParameter.gridSize.x), Random.Range(0, gridParameter.gridSize.y)].position;
gridParameter.targetState = grid[0, 0].position;
if (gridParameter.targetState.Equals(gridParameter.startState))
{
while (gridParameter.targetState.Equals(gridParameter.startState))
{
gridParameter.targetState = grid[Random.Range(0, gridParameter.gridSize.x), Random.Range(0, gridParameter.gridSize.y)].position;
}
}
// Setup du visuel du end state
grid[gridParameter.targetState.x, gridParameter.targetState.y].visual.GetComponent <Renderer>().material.color = Color.green;
grid[gridParameter.targetState.x, gridParameter.targetState.y].state = GridCell.GridState.End;
// Set du reward du end state
grid[gridParameter.targetState.x, gridParameter.targetState.y].r = 0.0f;
grid[gridParameter.targetState.x, gridParameter.targetState.y].v = 1000.0f;
// Reset des action possible au end state, on est a la fin donc on ne fait plus rien !
actionsDic[gridParameter.targetState] = new List <Action>();
// Creation des obstacles
for (int i = 0; i < gridParameter.unwalkableCellCount; i++)
{
Vector2Int idx = Vector2Int.zero;
do
{
idx = new Vector2Int(Random.Range(0, gridParameter.gridSize.x), Random.Range(0, gridParameter.gridSize.y));
} while (grid[idx.x, idx.y].state.Equals(GridCell.GridState.End) ||
grid[idx.x, idx.y].state.Equals(GridCell.GridState.Start) ||
ThereIsAnOtherObstacleSoClose(idx));
grid[idx.x, idx.y].state = GridCell.GridState.Bloc;
grid[idx.x, idx.y].r = 0.0f;
grid[idx.x, idx.y].v = 0.0f;
grid[idx.x, idx.y].visual.GetComponent <Renderer>().material.color = Color.black;
// On remove les actions possible sur les bloc
actionsDic[grid[idx.x, idx.y].position] = new List <Action>();
}
// Definition des actions possible pour chaque etat
for (int i = 0; i < gridParameter.gridSize.x; i++)
{
for (int j = 0; j < gridParameter.gridSize.y; j++)
{
if (grid[i, j].state.Equals(GridCell.GridState.End) ||
grid[i, j].state.Equals(GridCell.GridState.Bloc))
{
continue;
}
// Initialisation des actions possible pour le state [i, j]
foreach (var act in actions)
{
MoveAction a = act as MoveAction;
if (a.OutOfGridBound(grid[i, j].position + a.direction, ref gridParameter))
{
continue;
}
var nxt = grid[i, j].position + a.direction;
if (grid[nxt.x, nxt.y].state.Equals(GridCell.GridState.Bloc))
{
continue;
}
actionsDic[grid[i, j].position].Add(act);
}
}
}
// Setup de la camera afin d'avoir une vu global de la grid
cam.transform.position = new Vector3(gridParameter.gridSize.x * 0.5f, gridParameter.gridSize.y * 0.5f, -5);
cam.orthographic = true;
cam.orthographicSize = ((gridParameter.gridSize.x + gridParameter.gridSize.y) * 0.5f + 5f) * 0.5f;
// Initialisation de l'agent
this.agentGridWorld.Init(actionsDic, gridParameter.startState, gridParameter.targetState, theta: this.theta, gamma: gamma);
this.agentGridWorld.upArrow = this.upArrow;
this.agentGridWorld.rightArrow = this.rightArrow;
this.agentGridWorld.downArrow = this.downArrow;
this.agentGridWorld.leftArrow = this.leftArrow;
// Creation du visuel de l'agent
this.agentGridWorld.visual = Instantiate(agentPrefab, new Vector3(gridParameter.startState.x, gridParameter.startState.y, 0.0f), Quaternion.identity);
// Lancement du Reinforcement Learning
if (algo == GridWorld_Algo.POLICY_ITERATION)
{
StartCoroutine(UpdateWorldWithPolicyIteration());
}
else
{
this.agentGridWorld.ValueIteration(ref grid);
}
}
