// Sarsa thread
private void SarsaThread()
{
int iteration = 0;
TabuSearchExploration tabuPolicy = (TabuSearchExploration)sarsa.ExplorationPolicy;
EpsilonGreedyExploration explorationPolicy = (EpsilonGreedyExploration)tabuPolicy.BasePolicy;
while ((!needToStop) && (iteration < learningIterations))
{
explorationPolicy.Epsilon = explorationRate - ((double)iteration / learningIterations) * explorationRate;
sarsa.LearningRate = learningRate - ((double)iteration / learningIterations) * learningRate;
tabuPolicy.ResetTabuList();
var agentCurrentX = agentStartX;
var agentCurrentY = agentStartY;
int steps = 1;
int previousState = GetStateNumber(agentCurrentX, agentCurrentY);
int previousAction = sarsa.GetAction(previousState);
double reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, previousAction);
while ((!needToStop) && ((agentCurrentX != agentStopX) || (agentCurrentY != agentStopY)))
{
steps++;
tabuPolicy.SetTabuAction((previousAction + 2) % 4, 1);
int nextState = GetStateNumber(agentCurrentX, agentCurrentY);
int nextAction = sarsa.GetAction(nextState);
sarsa.UpdateState(previousState, previousAction, reward, nextState, nextAction);
reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, nextAction);
previousState = nextState;
previousAction = nextAction;
}
if (!needToStop)
{
sarsa.UpdateState(previousState, previousAction, reward);
}
System.Diagnostics.Debug.WriteLine(steps);
iteration++;
SetText(iterationBox, iteration.ToString());
}
// enable settings controls
EnableControls(true);
}
// Sarsa thread
private void SarsaThread()
{
int iteration = 0;
// curent coordinates of the agent
int agentCurrentX, agentCurrentY;
// exploration policy
TabuSearchExploration tabuPolicy = (TabuSearchExploration)sarsa.ExplorationPolicy;
EpsilonGreedyExploration explorationPolicy = (EpsilonGreedyExploration)tabuPolicy.BasePolicy;
// loop
while ((!needToStop) && (iteration < learningIterations))
{
// set exploration rate for this iteration
explorationPolicy.Epsilon = explorationRate - ((double)iteration / learningIterations) * explorationRate;
// set learning rate for this iteration
sarsa.LearningRate = learningRate - ((double)iteration / learningIterations) * learningRate;
// clear tabu list
tabuPolicy.ResetTabuList();
// reset agent's coordinates to the starting position
agentCurrentX = agentStartX;
agentCurrentY = agentStartY;
// steps performed by agent to get to the goal
int steps = 1;
// previous state and action
int previousState = GetStateNumber(agentCurrentX, agentCurrentY);
int previousAction = sarsa.GetAction(previousState);
// update agent's current position and get his reward
double reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, previousAction);
while ((!needToStop) && ((agentCurrentX != agentStopX) || (agentCurrentY != agentStopY)))
{
steps++;
// set tabu action
tabuPolicy.SetTabuAction((previousAction + 2) % 4, 1);
// get agent's next state
int nextState = GetStateNumber(agentCurrentX, agentCurrentY);
// get agent's next action
int nextAction = sarsa.GetAction(nextState);
// do learning of the agent - update his Q-function
sarsa.UpdateState(previousState, previousAction, reward, nextState, nextAction);
// update agent's new position and get his reward
reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, nextAction);
previousState = nextState;
previousAction = nextAction;
}
if (!needToStop)
{
// update Q-function if terminal state was reached
sarsa.UpdateState(previousState, previousAction, reward);
}
System.Diagnostics.Debug.WriteLine(steps);
iteration++;
// show current iteration
SetText(iterationBox, iteration.ToString());
}
// enable settings controls
EnableControls(true);
}
public void learn_test()
{
#region doc_main
// Fix the random number generator
Accord.Math.Random.Generator.Seed = 0;
// In this example, we will be using the Sarsa algorithm
// to make a robot learn how to navigate a map. The map
// is shown below, where a 1 denotes a wall and 0 denotes
// areas where the robot can navigate:
//
int[,] map =
{
{ 1, 1, 1, 1, 1, 1, 1, 1, 1 },
{ 1, 1, 0, 0, 0, 0, 0, 0, 1 },
{ 1, 1, 0, 0, 0, 1, 1, 0, 1 },
{ 1, 0, 0, 1, 0, 0, 0, 0, 1 },
{ 1, 0, 0, 1, 1, 1, 1, 0, 1 },
{ 1, 0, 0, 1, 1, 0, 0, 0, 1 },
{ 1, 1, 0, 1, 0, 0, 0, 0, 1 },
{ 1, 1, 0, 1, 0, 1, 1, 0, 1 },
{ 1, 1, 1, 1, 1, 1, 1, 1, 1 },
};
// Now, we define the initial and target points from which the
// robot will be spawn and where it should go, respectively:
int agentStartX = 1;
int agentStartY = 4;
int agentStopX = 7;
int agentStopY = 4;
// The robot is able to sense the environment though 8 sensors
// that capture whether the robot is near a wall or not. Based
// on the robot's current location, the sensors will return an
// integer number representing which sensors have detected walls
Func <int, int, int> getState = (int x, int y) =>
{
int c1 = (map[y - 1, x - 1] != 0) ? 1 : 0;
int c2 = (map[y - 1, x + 0] != 0) ? 1 : 0;
int c3 = (map[y - 1, x + 1] != 0) ? 1 : 0;
int c4 = (map[y + 0, x + 1] != 0) ? 1 : 0;
int c5 = (map[y + 1, x + 1] != 0) ? 1 : 0;
int c6 = (map[y + 1, x + 0] != 0) ? 1 : 0;
int c7 = (map[y + 1, x - 1] != 0) ? 1 : 0;
int c8 = (map[y + 0, x - 1] != 0) ? 1 : 0;
return(c1 | (c2 << 1) | (c3 << 2) | (c4 << 3) | (c5 << 4) | (c6 << 5) | (c7 << 6) | (c8 << 7));
};
// The actions are the possible directions the robot can go:
//
// - case 0: go to north (up)
// - case 1: go to east (right)
// - case 2: go to south (down)
// - case 3: go to west (left)
//
int learningIterations = 1000;
double explorationRate = 0.5;
double learningRate = 0.5;
double moveReward = 0;
double wallReward = -1;
double goalReward = 1;
// The function below specifies how the robot should perform an action given its
// current position and an action number. This will cause the robot to update its
// current X and Y locations given the direction (above) it was instructed to go:
Func <int, int, int, Tuple <double, int, int> > doAction = (int currentX, int currentY, int action) =>
{
// default reward is equal to moving reward
double reward = moveReward;
// moving direction
int dx = 0, dy = 0;
switch (action)
{
case 0: // go to north (up)
dy = -1;
break;
case 1: // go to east (right)
dx = 1;
break;
case 2: // go to south (down)
dy = 1;
break;
case 3: // go to west (left)
dx = -1;
break;
}
int newX = currentX + dx;
int newY = currentY + dy;
// check new agent's coordinates
if ((map[newY, newX] != 0) || (newX < 0) || (newX >= map.Columns()) || (newY < 0) || (newY >= map.Rows()))
{
// we found a wall or got outside of the world
reward = wallReward;
}
else
{
currentX = newX;
currentY = newY;
// check if we found the goal
if ((currentX == agentStopX) && (currentY == agentStopY))
{
reward = goalReward;
}
}
return(Tuple.Create(reward, currentX, currentY));
};
// After defining all those functions, we create a new Sarsa algorithm:
var explorationPolicy = new EpsilonGreedyExploration(explorationRate);
var tabuPolicy = new TabuSearchExploration(4, explorationPolicy);
var sarsa = new Sarsa(256, 4, tabuPolicy);
// curent coordinates of the agent
int agentCurrentX = -1;
int agentCurrentY = -1;
bool needToStop = false;
int iteration = 0;
// loop
while ((!needToStop) && (iteration < learningIterations))
{
// set exploration rate for this iteration
explorationPolicy.Epsilon = explorationRate - ((double)iteration / learningIterations) * explorationRate;
// set learning rate for this iteration
sarsa.LearningRate = learningRate - ((double)iteration / learningIterations) * learningRate;
// clear tabu list
tabuPolicy.ResetTabuList();
// reset agent's coordinates to the starting position
agentCurrentX = agentStartX;
agentCurrentY = agentStartY;
// steps performed by agent to get to the goal
int steps = 1;
// previous state and action
int previousState = getState(agentCurrentX, agentCurrentY);
int previousAction = sarsa.GetAction(previousState);
// update agent's current position and get his reward
var r = doAction(agentCurrentX, agentCurrentY, previousAction);
double reward = r.Item1;
agentCurrentX = r.Item2;
agentCurrentY = r.Item3;
while ((!needToStop) && ((agentCurrentX != agentStopX) || (agentCurrentY != agentStopY)))
{
steps++;
// set tabu action
tabuPolicy.SetTabuAction((previousAction + 2) % 4, 1);
// get agent's next state
int nextState = getState(agentCurrentX, agentCurrentY);
// get agent's next action
int nextAction = sarsa.GetAction(nextState);
// do learning of the agent - update his Q-function
sarsa.UpdateState(previousState, previousAction, reward, nextState, nextAction);
// update agent's new position and get his reward
r = doAction(agentCurrentX, agentCurrentY, nextAction);
reward = r.Item1;
agentCurrentX = r.Item2;
agentCurrentY = r.Item3;
previousState = nextState;
previousAction = nextAction;
}
if (!needToStop)
{
// update Q-function if terminal state was reached
sarsa.UpdateState(previousState, previousAction, reward);
}
iteration++;
}
// The end position for the robot will be (7, 4):
int finalPosX = agentCurrentX; // 7
int finalPosY = agentCurrentY; // 4;
#endregion
Assert.AreEqual(7, finalPosX);
Assert.AreEqual(4, finalPosY);
}
// Sarsa thread
private void SarsaThread()
{
int iteration = 0;
// exploration policy
TabuSearchExploration tabuPolicy = (TabuSearchExploration)_sarsa.ExplorationPolicy;
EpsilonGreedyExploration explorationPolicy = (EpsilonGreedyExploration)tabuPolicy.BasePolicy;
// loop
while ((!_needToStop) && (iteration < learningIterations))
{
// set exploration rate for this iteration
explorationPolicy.Epsilon = explorationRate - ((double)iteration / learningIterations) * explorationRate;
// set learning rate for this iteration
_sarsa.LearningRate = learningRate - ((double)iteration / learningIterations) * learningRate;
// clear tabu list
tabuPolicy.ResetTabuList();
// reset agent's coordinates to the starting position
_agentCurrX = _agentStartX;
_agentCurrY = _agentStartY;
// steps performed by agent to get to the goal
int steps = 1;
// previous state and action
int previousState = GetStateNumber(_agentCurrX, _agentCurrY);
int previousAction = _sarsa.GetAction(previousState);
// update agent's current position and get his reward
double reward = UpdateAgentPosition(previousAction);
while ((!_needToStop) && ((_agentCurrX != _agentStopX) || (_agentCurrY != _agentStopY)))
{
steps++;
// set tabu action
tabuPolicy.SetTabuAction((previousAction + 2) % 4, 1);
// get agent's next state
int nextState = GetStateNumber(_agentCurrX, _agentCurrY);
// get agent's next action
int nextAction = _sarsa.GetAction(nextState);
// do learning of the agent - update his Q-function
_sarsa.UpdateState(previousState, previousAction, reward, nextState, nextAction);
// update agent's new position and get his reward
reward = UpdateAgentPosition(nextAction);
previousState = nextState;
previousAction = nextAction;
}
if (!_needToStop)
{
// update Q-function if terminal state was reached
_sarsa.UpdateState(previousState, previousAction, reward);
}
iteration++;
Debug.Log(string.Format("{0} steps needed for iteration {1}.", steps, iteration));
}
_enableControls = true;
Debug.Log("SARSA training finished. Try to execute the solution.");
}
private void SarsaThread()
{
int iteration = 0;
// 当前坐标的代理 curent coordinates of the agent
int agentCurrentX, agentCurrentY;
// 探索策略 exploration policy
TabuSearchExploration tabuPolicy = (TabuSearchExploration)sarsa.ExplorationPolicy;
EpsilonGreedyExploration explorationPolicy = (EpsilonGreedyExploration)tabuPolicy.BasePolicy;
// loop
while ((!needToStop) && (iteration < learningIterations))
{
// 为这个迭代设置勘探速率 set exploration rate for this iteration
explorationPolicy.Epsilon = explorationRate - ((double)iteration / learningIterations) * explorationRate;
// 为迭代设置学习率 set learning rate for this iteration
sarsa.LearningRate = learningRate - ((double)iteration / learningIterations) * learningRate;
// 清除tabu列表 clear tabu list
tabuPolicy.ResetTabuList();
// 复位代理的坐标到起始位置 reset agent's coordinates to the starting position
agentCurrentX = _agentStartX;
agentCurrentY = _agentStartY;
// 代理执行的步骤以达到目标 steps performed by agent to get to the goal
int steps = 1;
// 以前的状态和动作 previous state and action
int previousState = GetStateNumber(agentCurrentX, agentCurrentY);
int previousAction = sarsa.GetAction(previousState);
// 更新代理的当前位置并得到他的奖励 update agent's current position and get his reward
double reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, previousAction);
while ((!needToStop) && ((agentCurrentX != _agentStopX) || (agentCurrentY != _agentStopY)))
{
steps++;
// 设置禁忌动作 set tabu action
tabuPolicy.SetTabuAction((previousAction + 2) % 4, 1);
// 获取代理的下一个状态 get agent's next state
int nextState = GetStateNumber(agentCurrentX, agentCurrentY);
// 获取代理的下一个动作 get agent's next action
int nextAction = sarsa.GetAction(nextState);
// 做学习代理 - 更新他的Q函数 do learning of the agent - update his Q-function
sarsa.UpdateState(previousState, previousAction, reward, nextState, nextAction);
// 更新代理的新位置并得到他的奖励 update agent's new position and get his reward
reward = UpdateAgentPosition(ref agentCurrentX, ref agentCurrentY, nextAction);
previousState = nextState;
previousAction = nextAction;
}
if (!needToStop)
{
// 如果达到终端状态，则更新Q函数 update Q-function if terminal state was reached
sarsa.UpdateState(previousState, previousAction, reward);
}
System.Diagnostics.Debug.WriteLine(steps);
iteration++;
// show current iteration
SetText(iterationBox, iteration.ToString());
}
// enable settings controls
EnableControls(true);
}