# Differences

This shows you the differences between two versions of the page.

 — blog:2019:0305_prob_qtable_learning [2019/03/05 17:59] (current) Line 1: Line 1: + ====== Probabilistic QTable learning ====== + + {{tag>​deep_learning}} + + As mentioned in my previous post on this subject, I feel it could be worth investigating a bit further on this "Q table learning"​ algorithm. + And more specifically,​ I want to try to introduce some kind of probabilistic action management in the system... Even if I'm not sure yet what kind of results this will give me. Let's just try and see ;-) + + ====== ====== + + ===== Initial model ===== + + So I spent some time trying to build a probabilistic Q table learning mechanism and ended with this code: ​def train_frozenlake_v2(numEpisodes,​ drate=7.0, stepCost=0.0,​ learningRate=0.8,​ eps=1.0, + max_epsilon=1.0,​ min_epsilon=0.001):​ + logDEBUG("​Building FrozenLake environment..."​) + env = gym.make('​FrozenLake-v0'​) + + #Initialize table with all zeros + nstates = env.observation_space.n + nactions = env.action_space.n + + # We don't want to start with a zero value Q table: + # We are going to consider those values as "​probability of action"​ on each row + # of the table, so they should represent correct probabilities:​ + Q = np.ones([nstates,​nactions])/​float(nactions) + + logDEBUG("​Qtable shape: %s" % str(Q.shape)) + + # Set learning parameters + lr = learningRate ​ # learning rate + y = .95  # gamma (ie. discount rate) + + # numEpisodes = 2000 + maxSteps = 99 + + # Exploration parameters + epsilon = 1.0                 # Exploration rate + # decay_rate = drate/​numEpisodes ​      # Exponential decay rate for exploration prob + decay_rate = drate       # Exponential decay rate for exploration prob + # decay rate detail: after numEpisodes episodes we reach a prob of exp(-7)~0.0009 + + actions = np.arange(nactions) + + # Array containing the total reward and number of steps per episode: + rList = np.zeros((numEpisodes,​ 2)) + + for i in range(numEpisodes):​ + #Reset environment and get first new observation + state = env.reset() + ​ + totalReward = 0 + done = False + step = 0 + + # logDEBUG("​Performing episode %d/​%d..."​ % (i, numEpisodes)) + path = np.zeros((maxSteps,​ 4))  # cols: state, action, prob, reward + + #The Q-Table learning algorithm + while step<​maxSteps:​ + # Check if we should do exploration or explotation:​ + exp_thres = random.uniform(0,​ 1) + + action = None + if exp_thres < epsilon: + # We do exploration:​ + action = env.action_space.sample() + # logDEBUG("​Using random action: %d" % action) + else: + # We do explotation,​ so we use our current Qtable + # but here we should consider a probability sampling instead of an ordinary argmax: + # action = np.argmax(Q[state,:​]) + action = np.random.choice(actions,​ size=None, replace=False,​ p=Q[state,:​]) + + # Get new state and reward from environment + newState,​reward,​done,​info = env.step(action) + + # Update Q-Table with new knowledge using Bellman formula: ​ + # No, we don't want to update the Qtable right now, + # instead we keep track of the action we selected for that state, ​ + # And we will decide later if this was a good choice or not and apply + # the reward only at that time: + path[step] = [float(state),​ float(action),​ Q[state, action], reward - stepCost] + + # Q[state,​action] = Q[state,​action] + lr*(reward - stepCost + y*np.max(Q[newState,:​]) - Q[state,​action]) + + # Update total reward: + totalReward += (reward - stepCost) + + # Update state: + state = newState + + step+=1 + + # Stop if we are done: + if done == True: + break + + ​ + + # When done with the episode we apply the training: + # We have to distribute the "total reward that we got", but we take into account the probability of the actions taken + # into the process: + ​ + # Remove the non-relevant data: + spath = path[:step] + + # coop ratio, the lower the "less cooperative distribution"​ we see: + weights = 1.0/​(eps+spath[:,​ 2]) + + # normalize the weights: + weights /= np.sum(weights) + # logDEBUG("​Normalized episode step weights: %s" % weights) + ​ + # If the total reward is 0.0 then we apply a small penalty: + # if totalReward == 0.0: + #     ​totalReward = -1.0 + + # Now we train the probability for each previously performed step: + Q[spath[:,​0].astype(int),​ spath[:,​1].astype(int)] += lr*totalReward*weights + + # Finally, we should renormalize our probability distributions:​ + Q = np.clip(Q, a_min=0.0, a_max=None) + + Q /= np.sum(Q, axis=1).reshape(-1,​ 1) + + # Then we reduce the exploration rate: + epsilon = min_epsilon + (max_epsilon - min_epsilon)*np.exp(-decay_rate*i) + + if (i+1)%1000==0:​ + # logDEBUG("​New epsilon value: %f" % epsilon) + # Compute the mean reward: + mvals = np.mean(rList[i-100:​i],​ axis=0) + # logDEBUG("​Mean reward: %f" % mvals[0]) + logDEBUG("​%d/​%d:​ Mean reward MA: %f" % (i+1, numEpisodes,​ mvals[0])) + # env.render() + + # And we assign our data for visualization:​ + rList[i] = [totalReward,​ step] + ​ + logDEBUG("​Final Qtable: %s" % np.around(Q,​6)) + + # return the data array: + return rList, Q​ + + * This was eventually producing "very sharp" Q tables, such as: <​code>​[[ 0.        0.        1.        0.      ] + [ 0.        0.        0.        1.      ] + [ 1.        0.        0.        0.      ] + [ 0.007265 ​ 0.492735 ​ 0.492735 ​ 0.007265] + [ 0.        0.        1.        0.      ] + [ 0.25      0.25      0.25      0.25    ] + [ 0.        0.        1.        0.      ] + [ 0.25      0.25      0.25      0.25    ] + [ 0.        1.        0.        0.      ] + [ 0.        1.        0.        0.      ] + [ 0.        1.        0.        0.      ] + [ 0.25      0.25      0.25      0.25    ] + [ 0.25      0.25      0.25      0.25    ] + [ 0.        0.        0.        1.      ] + [ 0.        1.        0.        0.      ] + [ 0.25      0.25      0.25      0.25    ]]​ + + * **But** then the results I could get out of this Qtable were not any better than the previous implementation. So I was wondering how this could be possible because I thought the environment was deterministic... But I was wrong, since the environment description contains this indication: + + <​note>​However,​ the ice is slippery, so you won't always move in the direction you intend. The surface is described using a grid like the following​ + + * **Stupid you Manu** ;-) You just have read this more carefully from the beginning! + + ===== Re-considering our implementation ===== + + * So basically, our actions in the environment will have a non-deterministic result. So let's try with a **non-slippery version of the environment** first. + ​ + * => And indeed, with a non-slippery environment,​ we quickly get to the best output: + + {{ projects:​nervseed:​qtable:​8-no_slip-reward.png?​1000 }} + + * But then, we can achieve the same kind of result with the bellman formula in such an environment:​ + + {{ projects:​nervseed:​qtable:​9-no_slip_bellman-reward.png?​1000 }} + + * **Note**: The "​slippery"​ behavior is defined as follow: + * if slippery: It would take 3 sub-steps and each step with 33.33% probability with each of the 3 adjacent directions + * if not slippery: only take 1 step and with 100% probability on this direction + + * So, the question is: is there something better we could do to training our Qtable given this new information ? Or should I just stick to that Bellman formula ? + + * Actually, if we think about it: when we get a reward after doing action $a_t$ in state $S_t$, we update the **quality** of that action/​state pair with something like: $Q(S_t, a_t) = Q(S_t, a_t) + \alpha*(R_t + \gamma \times max{ Q'​(S_{t+1},​a'​) } - Q(S_t,​a_t))$ + + * But, in this environment,​ when we reach a reward, we are at the end of the episode. So we need to wait for the next episode to see the new information we got, and that information is only available "one step before"​ the reward step from the previous episode: so we have some delay in the training updates. + + * => Why not instead try to "​propagate the reward backward"​ in the previous Qvalues that led to the rewarding state ? That might be worth trying at some point. + + * And now I'm thinking about something else: in the equation above, we use $max Q'​(S_{t+1},​a'​)$, where $S_{t+1}$ is the "next step" we are reaching with this action we just took. What if instead we could build a transition probability matrix ? ie. for each (S,a) pair we would build a probability to transition to another state S': so when it's time to update the Q(S,a) value then we could take this transition matrix into account and weight all the reachable new states accordingly ? => **let'​s give this a try** + + * And well, I think this was a good move actually, using this "​transition matrix"​ and reducing the learning rate to 0.1 I could achieve this kind of results: + + {{ projects:​nervseed:​qtable:​10-transition_mat-reward.png?​1000 }} + + And this is quite better than the reference implementation results below: + + {{ projects:​nervseed:​qtable:​11-reference_impl-reward.png?​1000 }} + + * So, the current implementation I'm using now (with transition matrix) is as follow: ​def train_frozenlake_v3(numEpisodes,​ drate=7.0, stepCost=0.0,​ learningRate=0.8,​ ename="​FrozenLake-v0",​ lrDecay=0.0001):​ + logDEBUG("​Building FrozenLake environment..."​) + env = gym.make(ename) + + #Initialize table with all zeros + nstates = env.observation_space.n + nactions = env.action_space.n + Q = np.zeros([nstates,​nactions]) + logDEBUG("​Qtable shape: %s" % str(Q.shape)) + + # Set learning parameters + lr = learningRate ​ # learning rate + y = .95  # gamma (ie. discount rate) + + # numEpisodes = 2000 + maxSteps = 99 + + # Exploration parameters + epsilon = 1.0                 # Exploration rate + max_epsilon = 1.0             # Exploration probability at start + min_epsilon = 0.01            # Minimum exploration probability ​ + # decay_rate = drate/​numEpisodes ​      # Exponential decay rate for exploration prob + decay_rate = drate       # Exponential decay rate for exploration prob + # decay rate detail: after numEpisodes episodes we reach a prob of exp(-7)~0.0009 + + # Array containing the total reward and number of steps per episode: + rList = np.zeros((numEpisodes,​ 2)) + + # Here we will also use a transition tensor: + trans = np.zeros((nactions,​ nstates, nstates)) + + for i in range(numEpisodes):​ + #Reset environment and get first new observation + state = env.reset() + ​ + totalReward = 0 + done = False + step = 0 + + # logDEBUG("​Performing episode %d/​%d..."​ % (i, numEpisodes)) + steps = [] + + #The Q-Table learning algorithm + while step<​maxSteps:​ + step+=1 + + # Check if we should do exploration or explotation:​ + exp_thres = random.uniform(0,​ 1) + + action = None + if exp_thres < epsilon: + # We do exploration:​ + action = env.action_space.sample() + # logDEBUG("​Using random action: %d" % action) + else: + # We do explotation,​ so we use our current Qtable: + action = np.argmax(Q[state,:​]) + + # logDEBUG("​State %d: performing action %d" % (state, action)) + + # Get new state and reward from environment + newState,​reward,​done,​info = env.step(action) + + # We have done one step, so we update the transition matrix: + trans[action,​ state, newState] += 1.0 + + # get the probabilistic next state weights: + weights = trans[action,​ state, :] / np.sum(trans[action,​ state, :]) + maxQs = np.max(Q, axis=1) ​ # this is the current max Q for each state. + # logDEBUG("​weights:​ %s" % weights) + # logDEBUG("​mawQs:​ %s" % maxQs) + + # compute the dot product of the maxQ and weights: + maxQ = np.dot(weights,​maxQs) + # logDEBUG("​MaxQ:​ %s" % maxQ) + + # Update Q-Table with new knowledge using Bellman formula: ​ + # Q[state,​action] = Q[state,​action] + lr*(reward - stepCost + y*np.max(Q[newState,:​]) - Q[state,​action]) + Q[state,​action] = Q[state,​action] + np.exp(-lrDecay*i)*lr*(reward - stepCost + y*maxQ - Q[state,​action]) + + # Update total reward: + totalReward += (reward - stepCost) + + # Update state: + state = newState + + # Stop if we are done: + if done == True: + break + ​ + # if totalReward > 0.0: + #     ​logDEBUG("​Total reward: %f" % totalReward) + + # Then we reduce the exploration rate: + epsilon = min_epsilon + (max_epsilon - min_epsilon)*np.exp(-decay_rate*i) + + if (i+1)%1000==0:​ + # logDEBUG("​New epsilon value: %f" % epsilon) + # logDEBUG("​results:​ %s" % rList[i-100:​i]) + + # Compute the mean reward: + mvals = np.mean(rList[i-100:​i],​ axis=0) + # logDEBUG("​Mean reward: %f" % mvals[0]) + logDEBUG("​%d/​%d:​ Mean reward: %f" % (i+1, numEpisodes,​ mvals[0])) + # env.render() + + # And we assign our data for visualization:​ + rList[i] = [totalReward,​ step] + ​ + logDEBUG("​Final Qtable: %s" % Q) + + # return the data array: + return rList, Q​ + + * **Note**: Actually we could also **"​pre-fill"​** the transition matrix with some **random sampling** to get less fluctuating initial probabilities. + + * Now, another step I think might also make a difference in the training is if we normalize the Q table for each row: this could given more "​weight"​ to the training changes, which might be a good thing ? **Trying that...** => Well, **no**: this doesn'​t help at all :-) I guess this is because we are messing the computation of the value of a given "next state" relative to the other possible next states. Never mind ;-) + + + ===== Recursive reward distribution ===== + + * When we get a reward at a given step, currently this will only chance the value of the immediate (state,​action) pair that lead to this reward. But why not consider instead that each previous step is a bit responsible for this success ? We could use our transition matrix again to figure out how to back propagate the reward value... Let's think about this for a moment. + + * SO I tried the following code: ​def propagateQValue(Q,​ trans, state, reward, y): + if abs(reward) <= 1e-6: + return Q + ​ + # logDEBUG("​Propagating reward %f in state %d" % (reward, state)) + + weights = trans[: , :, state] / (np.sum(np.sum(trans[:,​ :, state])) + 1e-10) + + # the weights above are given by action and then state, so we should transpose the matrix: + weights = np.transpose(weights) + + Q += reward * weights + + # We sum the weights for each state: + weights = np.sum(weights,​ axis=1) * y * reward + + for i in range(weights.shape[0]):​ + if i != state: + Q = propagateQValue(Q,​ trans, i, weights[i], y) + ​ + return Q + + def train_frozenlake_v4(numEpisodes,​ drate=7.0, stepCost=0.0,​ learningRate=0.8,​ ename="​FrozenLake-v0",​ lrDecay=0.0001):​ + logDEBUG("​Building FrozenLake environment..."​) + env = gym.make(ename) + + #Initialize table with all zeros + nstates = env.observation_space.n + nactions = env.action_space.n + Q = np.zeros([nstates,​nactions]) + logDEBUG("​Qtable shape: %s" % str(Q.shape)) + + # Set learning parameters + lr = learningRate ​ # learning rate + y = .95  # gamma (ie. discount rate) + + # numEpisodes = 2000 + maxSteps = 99 + + # Exploration parameters + epsilon = 1.0                 # Exploration rate + max_epsilon = 1.0             # Exploration probability at start + min_epsilon = 0.01            # Minimum exploration probability ​ + # decay_rate = drate/​numEpisodes ​      # Exponential decay rate for exploration prob + decay_rate = drate       # Exponential decay rate for exploration prob + # decay rate detail: after numEpisodes episodes we reach a prob of exp(-7)~0.0009 + + # Array containing the total reward and number of steps per episode: + rList = np.zeros((numEpisodes,​ 2)) + + # Here we will also use a transition tensor: + trans = np.zeros((nactions,​ nstates, nstates)) + + for i in range(numEpisodes):​ + #Reset environment and get first new observation + state = env.reset() + ​ + totalReward = 0 + done = False + step = 0 + + # logDEBUG("​Performing episode %d/​%d..."​ % (i, numEpisodes)) + steps = [] + + #The Q-Table learning algorithm + while step<​maxSteps:​ + step+=1 + + # Check if we should do exploration or explotation:​ + exp_thres = random.uniform(0,​ 1) + + action = None + if exp_thres < epsilon: + # We do exploration:​ + action = env.action_space.sample() + # logDEBUG("​Using random action: %d" % action) + else: + # We do explotation,​ so we use our current Qtable: + action = np.argmax(Q[state,:​]) + + # logDEBUG("​State %d: performing action %d" % (state, action)) + + # Get new state and reward from environment + newState,​reward,​done,​info = env.step(action) + + # We have done one step, so we update the transition matrix: + trans[action,​ state, newState] += 1.0 + + # get the probabilistic next state weights: + # weights = trans[action,​ state, :] / np.sum(trans[action,​ state, :]) + # maxQs = np.max(Q, axis=1) ​ # this is the current max Q for each state. + # logDEBUG("​weights:​ %s" % weights) + # logDEBUG("​mawQs:​ %s" % maxQs) + + # compute the dot product of the maxQ and weights: + # maxQ = np.dot(weights,​maxQs) + # logDEBUG("​MaxQ:​ %s" % maxQ) + + # Update Q-Table with new knowledge using Bellman formula: ​ + # Q[state,​action] = Q[state,​action] + np.exp(-lrDecay*i)*lr*(reward - stepCost + y*maxQ - Q[state,​action]) + + # if there is a reward to distribute (either positive or negative), we should consider how we reached this current (state/​action) + # And we should do so recursively:​ + Q = propagateQValue(Q,​ trans, newState, np.exp(-lrDecay*i)*lr*reward,​ y) + + # if reward != 0.0: + #     ​weights = trans[: , :, state] / np.sum(np.sum(trans[:,​ :, state])) + + #     # the weights above are given by action and then state, so we should transpose the matrix: + #     ​weights = np.transpose(weights) + + #     Q += np.exp(-lrDecay*i)*lr*reward*weights + + # Update total reward: + totalReward += (reward - stepCost) + + # Update state: + state = newState + + # Stop if we are done: + if done == True: + break + + # Then we reduce the exploration rate: + epsilon = min_epsilon + (max_epsilon - min_epsilon)*np.exp(-decay_rate*i) + + if (i+1)%1000==0:​ + # Compute the mean reward: + mvals = np.mean(rList[i-100:​i],​ axis=0) + # logDEBUG("​Mean reward: %f" % mvals[0]) + logDEBUG("​%d/​%d:​ Mean reward: %f" % (i+1, numEpisodes,​ mvals[0])) + + # And we assign our data for visualization:​ + rList[i] = [totalReward,​ step] + ​ + logDEBUG("​Final Qtable: %s" % Q) + + # return the data array: + return rList, Q​ + + * => Unfortunately,​ the results with this implementation are really bad (see image below), and it's also extremely slow due to the recursive call to **propagateQValue()**. So, not a good direction I guess :-). + + {{ projects:​nervseed:​qtable:​12-back_prop-reward.png?​1000 }} + + ===== What to learn from failed episodes ===== + + * One thing we could also take out of this experiment is that, in this environment,​ we only train our Q table with positive rewards. But what about when we fail at a given episode ? (ie. we fall in a hole...) In that case, we are not learning anything. Maybe it would be worth thinking about a mechanism to change the Qvalues when we face a negative episode ? + + * But well, I feel I spent enough time on this "Q table" learning for now, and I want to move to a neural network implementation now, so I'll just stop here for now and see later if there is a need for me to come back to it. +