s442720/Waiter_group
1
0
Fork 0
forked from s444519/Waiter_group
Code Releases Activity

Individual Project #2; s442720

Browse Source
This commit is contained in:
Tao-Sen Chang 2020-05-10 12:42:43 +00:00
parent f3ddde84e3
commit d9a5f26c02
1 changed files with 235 additions and 235 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
Reinforcement_learning.md
View File

@ -1,16 +1,16 @@
# Reinforcement learning for route planning in restaurant # Reinforcement learning for route planning in restaurant
##### Tao-Sen Chang s442720 ##### Tao-Sen Chang s442720
##### We did the route planning by special algorithm on last task. In this machine learning sub-project I try to show different apporach for the agent who can traversal multiple destinations on the grid system, and of course, get the shortest path of the route. The method is called reinforcement learning. ###### We did the route planning by special algorithm on last task. In this machine learning sub-project I try to show different approach for the agent who can traversal multiple destinations on the grid system, and of course, get the shortest path of the route. The method is called reinforcement learning.
## What is reinforcement learning? ## What is reinforcement learning?
##### Reinforcement learning is how software agents ought to take actions in an environment in order to maximize the notion of cumulative reward. The agent makes a sequence of decisions, and learn to perform best actions every step. For example, in my project there is a waiter in the grid and he have to reach many tables for serving the meal, so he must learn the shortest path to get a table. ###### Reinforcement learning is how software agents ought to take actions in an environment in order to maximize the notion of cumulative reward. The agent makes a sequence of decisions, and learn to perform the best actions every step. For example, in my project there is a waiter in the grid and he has to reach many tables for serving the meal, so he must learn the shortest path to get a table.
## How to do that? ## How to do that?
##### The idea of how to complete the reinforcement learning is not quite easy. However, there is an example - rat in a maze. Instead of writing the whole algorithm at beginning, I use the existing codes(tools) by adjusting some parameters to train the agent on our 16x16 grid. ###### The idea of how to complete the reinforcement learning is not quite easy. However, there is an example - rat in a maze. Instead of writing the whole algorithm at the beginning, I use the existing codes(tools) by adjusting some parameters to train the agent on our 16x16 grid.
https://www.samyzaf.com/ML/rl/qmaze.html https://www.samyzaf.com/ML/rl/qmaze.html
##### I train the agent(waiter) with rewards and penalties, the waiter in the above grid gets a small penalty for every legal move. The reason is that we want it to get to the target table in the shortest possible path. However, the shortest path to the target table is sometimes long and winding, and our agent (the waiter) may have to endure many errors until he gets to the table. ###### I train the agent(waiter) with rewards and penalties, the waiter in the above grid gets a small penalty for every legal move. The reason is that we want it to get to the target table in the shortest possible path. However, the shortest path to the target table is sometimes long and winding, and our agent (the waiter) may have to endure many errors until he gets to the table.
##### For example, one of the training parameters(rewards) are: ###### For example, one of the training parameters(rewards) are:
``` ```
if rat_row == win_target_x and rat_col == win_target_y: # if reach the final target if rat_row == win_target_x and rat_col == win_target_y: # if reach the final target
return 1.0 return 1.0
@ -29,11 +29,11 @@ if (rat_row, rat_col) in self.curr_win_targets: # if reach any table
self.min_reward = -0.5 * self.maze.size self.min_reward = -0.5 * self.maze.size
``` ```
## Q-learning ## Q-learning
##### We want to get the maximum reward from each action in a state. Here defines action=π(s). ###### We want to get the maximum reward from each action in a state. Here defines action=π(s).
##### Q(s,a) = the maximum total reward we can get by choosing action a in state s. Hence it's obvious that we get the function π(s)=argmaxQ(s,ai) Now the question is how to get Q(s,a)? ###### Q(s,a) = the maximum total reward we can get by choosing action a in state s. Hence it's obvious that we get the function π(s)=argmaxQ(s,ai) Now the question is how to get Q(s,a)?
##### There is a solution called Bellman's Equation: Q(s,a) = R(s,a) + maxQ(s′,ai) ###### There is a solution called Bellman's Equation: Q(s,a) = R(s,a) + maxQ(s′,ai)
##### R(s,a) is the reward in current state s, action a. And s′ means the next state, so maxQ(s′,ai) means the maximum reward in 4 actions from next state. In the code we have the Experience Class to memorize each "episode", but the memory is limited, therefore if reach the max_memory, then delete the old episode which has lower effect to current episode. ###### R(s,a) is the reward in current state s, action a. And s′ means the next state, so maxQ(s′,ai) means the maximum reward in 4 actions from next state. In the code we have the Experience Class to memorize each "episode", but the memory is limited, therefore if reach the max_memory, then delete the old episode which has lower effect to current episode.
##### There is a coefficient called discount factor, usually denoted by γ which is required for the Bellman equation for stochastic environments. So the new Bellman's Equation can be written as Q(s,a) = R(s,a) + γ * maxQ(s′,ai). This discount factor is to diminish the effects which are far from current state. ###### There is a coefficient called discount factor, usually denoted by γ which is required for the Bellman equation for stochastic environments. So the new Bellman's Equation can be written as Q(s,a) = R(s,a) + γ * maxQ(s′,ai). This discount factor is to diminish the effects which are far from current state.
``` ```
class Experience(object): class Experience(object):
def __init__(self, model, max_memory=100, discount=0.95): def __init__(self, model, max_memory=100, discount=0.95):
@ -76,8 +76,8 @@ class Experience(object):
``` ```
## Training ## Training
##### Following is the algorithm for training neural network model to solve the problem. One epoch means one loop of the training, and in each epoch the agent will finally become either "win" or "lose". ###### Following is the algorithm for training neural network model to solve the problem. One epoch means one loop of the training, and in each epoch the agent will finally become either "win" or "lose".
##### Another coefficient "epsilon" is exploration factor which decides the probability of whether the agent will perform new actions instead of following the previous experiences (which is called exploitation). By this way the agent could not only collect better rewards from previous experiences, but also have the chances to explore unknow area where might get more rewards. If one of the strategy is determined, then let's start training it by neural network. (inputs: size equals to the maze size, targets: size is the same as the number of actions (4 in our case)). ###### Another coefficient "epsilon" is exploration factor which decides the probability of whether the agent will perform new actions instead of following the previous experiences (which is called exploitation). By this way the agent could not only collect better rewards from previous experiences, but also have the chances to explore unknown area where might get more rewards. If one of the strategies is determined, then let's start training it by neural network. (inputs: size equals to the maze size, targets: size is the same as the number of actions (4 in our case)).
``` ```
# Exploration factor # Exploration factor
epsilon = 0.1 epsilon = 0.1
@ -182,7 +182,7 @@ def qtrain(model, maze, **opt):
``` ```
## Testing ## Testing
##### Use this algorithm to our 16x16 grid and train. ###### Use this algorithm to our 16x16 grid and train.
``` ```
grid = [[1 for x in range(16)] for y in range(16)] grid = [[1 for x in range(16)] for y in range(16)]
table1 = Table(2, 2) table1 = Table(2, 2)
@ -199,11 +199,11 @@ maze = np.array(grid)
model = build_model(maze) model = build_model(maze)
qtrain(model, maze, epochs=1000, max_memory=8*maze.size, data_size=32) qtrain(model, maze, epochs=1000, max_memory=8*maze.size, data_size=32)
``` ```
##### Also I create a list called win_targets to put the position of tables in the grid. ###### Also I create a list called win_targets to put the position of tables in the grid.
``` ```
win_targets = [(4, 4),(4, 9),(4, 14),(9, 4),(9, 9),(9, 14),(14, 4),(14, 9)] win_targets = [(4, 4),(4, 9),(4, 14),(9, 4),(9, 9),(9, 14),(14, 4),(14, 9)]
``` ```
##### After tons of trainings, I realize it is not an easy task to obtain the shortest route in every training - that means most of the trainings are failed - especially in the case that the win_targets has more targets. For example, the result of training 8 targets is like this(part of result): ###### After tons of training, I realize it is not an easy task to obtain the shortest route in every training - that means most of the training are failed - especially in the case that the win_targets has more targets. For example, the result of training 8 targets is like this(part of result):
``` ```
... ...
Epoch: 167/14999 | Loss: 0.0299 | Episodes: 407 | Win count: 63 | Win rate: 0.422 | time: 2.44 hours Epoch: 167/14999 | Loss: 0.0299 | Episodes: 407 | Win count: 63 | Win rate: 0.422 | time: 2.44 hours
@ -215,7 +215,7 @@ Epoch: 172/14999 | Loss: 0.0121 | Episodes: 362 | Win count: 67 | Win rate: 0.43
Epoch: 173/14999 | Loss: 0.0101 | Episodes: 484 | Win count: 68 | Win rate: 0.445 | time: 2.52 hours Epoch: 173/14999 | Loss: 0.0101 | Episodes: 484 | Win count: 68 | Win rate: 0.445 | time: 2.52 hours
... ...
``` ```
##### The only one which is successful contains 4 targets(win_targets = [(4, 4),(4, 9),(4, 14),(9, 4)]) ###### The only one which is successful contains 4 targets(win_targets = [(4, 4),(4, 9),(4, 14),(9, 4)])
``` ```
... ...
Epoch: 223/14999 | Loss: 0.0228 | Episodes: 30 | Win count: 165 | Win rate: 0.906 | time: 64.02 minutes Epoch: 223/14999 | Loss: 0.0228 | Episodes: 30 | Win count: 165 | Win rate: 0.906 | time: 64.02 minutes
@ -229,7 +229,7 @@ Epoch: 460/14999 | Loss: 0.0236 | Episodes: 60 | Win count: 401 | Win rate: 1.00
Reached 100% win rate at epoch: 460 Reached 100% win rate at epoch: 460
n_epoch: 460, max_mem: 2048, data: 32, time: 1.48 hours n_epoch: 460, max_mem: 2048, data: 32, time: 1.48 hours
``` ```
##### In my opinion, there are 3 reasons cause such bad results. ###### In my opinion, there are 3 reasons cause such bad results.
##### 1. The parameters in the algorithm are not optimal including the rewards, exploration rate, and discount factor. To adjust the parameters and to validate them costs lots of time, and the most intuitive way is always not the best solution. For example, the parameters of 4 targets are fine, but if the number of targets expanded to 8, the parameters are not just 1/2 of the original ones. ###### 1. The parameters in the algorithm are not optimal including the rewards, exploration rate, and discount factor. To adjust the parameters and to validate them costs lots of time, and the most intuitive way is always not the best solution. For example, the parameters of 4 targets are fine, but if the number of targets expanded to 8, the parameters are not just 1/2 of the original ones.
##### 2. Beacuse of the exploration rate, every time the same training and testing data may have different result. It increases the difficulty to verify our result. Only way to check whether the parameters generate ideal results is training continuously until we collect sufficient data. ###### 2. Because of the exploration rate, every time the same training and testing data may have a different result. It increases the difficulty to verify our result. The only way to check whether the parameters generate ideal results is training continuously until we collect sufficient data.
##### 3. The algorithm is for a rat in a maze at beginning, and the number of default target is only one. If we apply it for multiple targets, there may be inadequate for some reason. Moreover, the default size is 7x7 grid. It is possible that the 16x16 grid is too huge for this algorithm. ###### 3. The algorithm is for a rat in a maze at the beginning, and the number of default target is only one. If we apply it for multiple targets, there may be inadequate for some reason. Moreover, the default size is 7x7. It is possible that the 16x16 grid is too huge for this algorithm.