class SemiGradientTDControl:
def __init__(self,
nParams,
nActions,
alpha,
approximationFunctionArgs,
actionSelectionMethod="egreedy",
epsilon=0.01):
self.name = "Generic SemiGradient TD Control Class"
self.nParams = nParams
self.nActions = nActions
self.alpha = alpha
self.af_kwargs = approximationFunctionArgs
self.af = getValueFromDict(self.af_kwargs, "af")
self.afd = getValueFromDict(self.af_kwargs, "afd")
self.w = np.zeros([self.nParams], dtype=float)
self.policy = FunctionApproximationPolicy(
self.nParams,
self.nActions,
self.af_kwargs,
actionSelectionMethod=actionSelectionMethod,
epsilon=epsilon)
def selectAction(self, state):
return self.policy.selectAction(state)
def getValue(self, state, action=None):
if action is None:
return np.array([
self.af(self.w, state, a, **self.af_kwargs)
for a in range(self.nActions)
])
else:
return self.af(self.w, state, action, **self.af_kwargs)
def getName(self):
return self.name
def reset(self):
self.w = np.zeros([self.nParams], dtype=float)
self.policy.reset()
def getGreedyAction(self, state, actionsAvailable=None):
q = np.array([
self.af(self.w, state, a, **self.af_kwargs)
for a in range(self.nActions)
])
if (actionsAvailable is None):
actionValues = q[:]
actionList = np.array(range(self.nActions))
else:
actionValues = q[actionsAvailable]
actionList = np.array(actionsAvailable)
actionIdx = selectAction_greedy(actionValues)
return actionList[actionIdx]
class TrueOnlineSARSA:
def __init__(self,
nParams,
nActions,
alpha,
gamma,
lambd,
approximationFunctionArgs,
actionSelectionMethod="egreedy",
epsilon=0.01):
self.name = "True Online SARSA"
self.nParams = nParams
self.nActions = nActions
self.alpha = alpha
self.gamma = gamma
self.lambd = lambd
self.af_kwargs = approximationFunctionArgs
self.af = getValueFromDict(self.af_kwargs, "af")
self.ftf = getValueFromDict(self.af_kwargs, "ftf")
self.w = np.zeros([self.nParams], dtype=np.float)
self.z = np.zeros([self.nParams], dtype=np.float)
self.q_old = 0.0
self.policy = FunctionApproximationPolicy(
self.nParams,
self.nActions,
self.af_kwargs,
actionSelectionMethod=actionSelectionMethod,
epsilon=epsilon)
def update(self, episode):
t = len(episode) - 2
state = episode[t]["state"]
action = episode[t]["action"]
reward = episode[t + 1]["reward"]
next_state = episode[t + 1]["state"]
next_action = episode[t + 1]["action"]
done = episode[t + 1]["done"]
x = self.ftf(state, action, **self.af_kwargs)
xx = self.ftf(next_state, next_action, **self.af_kwargs)
q = self.getValue(state, action)
q_next = self.getValue(next_state, next_action)
td_error = reward + self.gamma * q_next - q
self.z = self.gamma * self.lambd * self.z + (
1 - self.alpha * self.gamma * self.lambd * np.dot(self.z, x)) * x
self.w += self.alpha * (td_error + q - self.q_old
) * self.z - self.alpha * (q - self.q_old) * x
self.policy.update(self.w)
self.q_old = q_next
if done:
self.z *= 0.0
self.q_old = 0.0
def getValue(self, state, action=None):
if action is None:
return np.array([
self.af(self.w, state, action, **self.af_kwargs)
for action in range(self.nActions)
])
else:
return self.af(self.w, state, action, **self.af_kwargs)
def selectAction(self, state):
return self.policy.selectAction(state)
def reset(self):
self.w = np.zeros([self.nParams], dtype=np.float)
self.z = np.zeros([self.nParams], dtype=np.float)
self.q_old = 0.0
def getName(self):
return self.name
class SARSALambda:
def __init__(self,
nParams,
nActions,
alpha,
gamma,
lambd,
approximationFunctionArgs,
doAccumulateTraces=False,
doClearTraces=False,
actionSelectionMethod="egreedy",
epsilon=0.01):
self.name = "SARSA(Lambda)"
self.nParams = nParams
self.nActions = nActions
self.alpha = alpha
self.gamma = gamma
self.lambd = lambd
self.af_kwargs = approximationFunctionArgs
self.af = getValueFromDict(self.af_kwargs, "af")
self.ftf = getValueFromDict(self.af_kwargs, "ftf")
self.doAccumulateTraces = doAccumulateTraces
self.doClearTraces = doClearTraces
self.w = np.zeros([self.nParams], dtype=np.float)
self.z = np.zeros([self.nParams], dtype=np.float)
self.policy = FunctionApproximationPolicy(
self.nParams,
self.nActions,
self.af_kwargs,
actionSelectionMethod=actionSelectionMethod,
epsilon=epsilon)
def update(self, episode):
t = len(episode) - 2
state = episode[t]["state"]
action = episode[t]["action"]
reward = episode[t + 1]["reward"]
next_state = episode[t + 1]["state"]
next_action = episode[t + 1]["action"]
done = episode[t + 1]["done"]
x = self.ftf(state, action, **self.af_kwargs)
xx = self.ftf(next_state, next_action, **self.af_kwargs)
td_error = reward
for i in np.nonzero(x)[0]:
td_error -= self.w[i]
if self.doAccumulateTraces:
self.z[i] += 1
else:
self.z[i] = 1
if done:
self.w += self.alpha * td_error * self.z
self.policy.update(self.w)
self.z *= 0.0
else:
for i in np.nonzero(xx)[0]:
td_error += self.gamma * self.w[i]
self.w += self.alpha * td_error * self.z
self.policy.update(self.w)
self.z = self.gamma * self.lambd * self.z
if self.doClearTraces:
idxToClear = np.array(np.ones(self.nParams), dtype=int)
idxToClear[np.nonzero(x)[0]] = 0
idxToClear[np.nonzero(xx)[0]] = 0
self.z[idxToClear] = 0.0
def getValue(self, state, action=None):
if action is None:
return np.array([
self.af(self.w, state, action, **self.af_kwargs)
for action in range(self.nActions)
])
else:
return self.af(self.w, state, action, **self.af_kwargs)
def selectAction(self, state):
return self.policy.selectAction(state)
def reset(self):
self.w = np.zeros([self.nParams], dtype=np.float)
self.z = np.zeros([self.nParams], dtype=np.float)
def getName(self):
return self.name