torch_training/Getting_Started_Training.md

-# How to train an Agent on Flatland
\ No newline at end of file
+# How to train an Agent on Flatland
+Quick introduction on how to train a simple DQN agent using Flatland and Pytorch. At the end of this Tutorial you should be able to train a single agent to navigate in Flatland.
+We use the `training_navigation.py` ([here](https://gitlab.aicrowd.com/flatland/baselines/blob/master/torch_training/training_navigation.py)) file to train a simple agent with the tree observation to solve the navigation task.
+
+## Actions in Flatland
+Flatland is a railway simulation. Thus the actions of an agent are strongly limited to the railway network. This means that in many cases not all actions are valid.
+The possible actions of an agent are
+
+- 0 *Do Nothing*:  If the agent is moving it continues moving, if it is stopped it stays stopped
+- 1 *Deviate Left*: This action is only valid at cells where the agent can change direction towards left. If action is chosen, the left transition and a rotation of the agent orientation to the left is executed. If the agent is stopped at any position, this action will cause it to start moving in any cell where forward or left is allowed!
+- 2 *Go Forward*: This action will start the agent when stopped. At switches this will chose the forward direction.
+- 3 *Deviate Right*: Exactly the same as deviate left but for right turns.
+- 4 *Stop*: This action causes the agent to stop, this is necessary to avoid conflicts in multi agent setups (Not needed for navigation).
+
+## Tree Observation
+Flatland offers three basic observations from the beginning. We encourage you to develop your own observations that are better suited for this specific task.
+
+For the navigation training we start with the Tree Observation as agents will learn the task very quickly using this observation.
+The tree observation exploits the fact that a railway network is a graph and thus the observation is only built along allowed transitions in the graph.
+
+Here is a small example of a railway network with an agent in the top left corner. The tree observation is build by following the allowed transitions for that agent.
+
+![Small_Network](https://i.imgur.com/utqMx08.png)
+
+As we move along the allowed transitions we build up a tree where a new node is created at every cell where the agent has different possibilities (Switch), dead-end or the target is reached.
+It is important to note that the tree observation is always build according to the orientation of the agent at a given node. This means that each node always has 4 branches coming from it in the directions *Left, Forward, Right and Backward*. These are illustrated with different colors in the figure below. The tree is build form the example rail above. Nodes where there are no possibilities are filled with `-inf` and are not all shown here for simplicity. The tree however, always has the same number of nodes for a given tree depth.
+
+![Tree_Observation](https://i.imgur.com/VsUQOQz.png)
+
+### Node Information
+Each node is filled with information gathered along the path to the node. Currently each node contains 9 features:
+
+- 1: if own target lies on the explored branch the current distance from the agent in number of cells is stored.
+
+- 2: if another agents target is detected the distance in number of cells from current agent position is stored.
+
+- 3: if another agent is detected the distance in number of cells from current agent position is stored.
+
+- 4: possible conflict detected (This only works when we use a predictor and will not be important in this tutorial)
+
+
+- 5: if an not usable switch (for agent) is detected we store the distance. An unusable switch is a switch where the agent does not have any choice of path, but other agents coming from different directions might. 
+
+
+- 6: This feature stores the distance (in number of cells) to the next node (e.g. switch or target or dead-end)
+
+- 7: minimum remaining travel distance from node to the agent's target given the direction of the agent if this path is chosen
+
+
+- 8: agent in the same direction found on path to node
+    - n = number of agents present same direction (possible future use: number of other agents in the same direction in this branch)
+    - 0 = no agent present same direction
+
+- 9: agent in the opposite direction on path to node
+    - n = number of agents present other direction than myself
+    - 0 = no agent present other direction than myself
+
+For training purposes the tree is flattend into a single array.
+
+## Training
+### Setting up the environment
+Before you get started with the training make sure that you have [pytorch](https://pytorch.org/get-started/locally/) installed.
+Let us now train a simPle double dueling DQN agent to navigate to its target on flatland. We start by importing flatland
+
+```
+from flatland.envs.generators import complex_rail_generator
+from flatland.envs.observations import TreeObsForRailEnv
+from flatland.envs.rail_env import RailEnv
+from flatland.utils.rendertools import RenderTool
+from utils.observation_utils import norm_obs_clip, split_tree
+```
+
+For this simple example we want to train on randomly generated levels using the `complex_rail_generator`. We use the following parameter for our first experiment:
+
+```
+# Parameters for the Environment
+x_dim = 10
+y_dim = 10
+n_agents = 1
+n_goals = 5
+min_dist = 5
+```
+
+As mentioned above, for this experiment we are going to use the tree observation and thus we load the observation builder:
+
+```
+# We are training an Agent using the Tree Observation with depth 2
+observation_builder = TreeObsForRailEnv(max_depth=2)
+```
+
+And pass it as an argument to the environment setup
+
+```
+env = RailEnv(width=x_dim,
+              height=y_dim,
+              rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
+                                                    max_dist=99999,
+                                                    seed=0),
+              obs_builder_object=observation_builder,
+              number_of_agents=n_agents)
+```
+
+We have no successfully set up the environment for training. To visualize it in the renderer we also initiate the renderer with.
+
+```
+env_renderer = RenderTool(env, gl="PILSVG", )
+```
+
+### Setting up the agent
+
+To set up a appropriate agent we need the state and action space sizes. From the discussion above about the tree observation we end up with:
+
+[**Adrian**: I just wonder, why this is not done in seperate method in the the observation: get_state_size, then we don't have to write down much more. And the user don't need to 
+understand anything about the observation. I suggest moving this into the observation, base ObservationBuilder declare it as an abstract method. ... ] 
+
+```
+# Given the depth of the tree observation and the number of features per node we get the following state_size
+features_per_node = 9
+tree_depth = 2
+nr_nodes = 0
+for i in range(tree_depth + 1):
+    nr_nodes += np.power(4, i)
+state_size = features_per_node * nr_nodes
+
+# The action space of flatland is 5 discrete actions
+action_size = 5
+```
+
+In the `training_navigation.py` file you will find further variable that we initiate in order to keep track of the training progress.
+Below you see an example code to train an agent. It is important to note that we reshape and normalize the tree observation provided by the environment to facilitate training.
+To do so, we use the utility functions `split_tree(tree=np.array(obs[a]), num_features_per_node=features_per_node, current_depth=0)` and `norm_obs_clip()`. Feel free to modify the normalization as you see fit.
+
+```
+# Split the observation tree into its parts and normalize the observation using the utility functions.
+    # Build agent specific local observation
+    for a in range(env.get_num_agents()):
+        rail_data, distance_data, agent_data = split_tree(tree=np.array(obs[a]),
+                                                          num_features_per_node=features_per_node,
+                                                          current_depth=0)
+        rail_data = norm_obs_clip(rail_data)
+        distance_data = norm_obs_clip(distance_data)
+        agent_data = np.clip(agent_data, -1, 1)
+        agent_obs[a] = np.concatenate((np.concatenate((rail_data, distance_data)), agent_data))
+```
+
+We now use the normalized `agent_obs` for our training loop:
+[**Adrian**: Same question as above, why not done in the observation class?]
+
+```
+for trials in range(1, n_trials + 1):
+
+    # Reset environment
+    obs, info = env.reset(True, True)
+    if not Training:
+        env_renderer.set_new_rail()
+
+    # Split the observation tree into its parts and normalize the observation using the utility functions.
+    # Build agent specific local observation
+    for a in range(env.get_num_agents()):
+        rail_data, distance_data, agent_data = split_tree(tree=np.array(obs[a]),
+                                                          num_features_per_node=features_per_node,
+                                                          current_depth=0)
+        rail_data = norm_obs_clip(rail_data)
+        distance_data = norm_obs_clip(distance_data)
+        agent_data = np.clip(agent_data, -1, 1)
+        agent_obs[a] = np.concatenate((np.concatenate((rail_data, distance_data)), agent_data))
+
+    # Reset score and done
+    score = 0
+    env_done = 0
+
+    # Run episode
+    for step in range(max_steps):
+
+        # Only render when not triaing
+        if not Training:
+            env_renderer.renderEnv(show=True, show_observations=True)
+
+        # Chose the actions
+        for a in range(env.get_num_agents()):
+            if not Training:
+                eps = 0
+
+            action = agent.act(agent_obs[a], eps=eps)
+            action_dict.update({a: action})
+
+            # Count number of actions takes for statistics
+            action_prob[action] += 1
+
+        # Environment step
+        next_obs, all_rewards, done, _ = env.step(action_dict)
+
+        for a in range(env.get_num_agents()):
+            rail_data, distance_data, agent_data = split_tree(tree=np.array(next_obs[a]),
+                                                              num_features_per_node=features_per_node,
+                                                              current_depth=0)
+            rail_data = norm_obs_clip(rail_data)
+            distance_data = norm_obs_clip(distance_data)
+            agent_data = np.clip(agent_data, -1, 1)
+            agent_next_obs[a] = np.concatenate((np.concatenate((rail_data, distance_data)), agent_data))
+
+        # Update replay buffer and train agent
+        for a in range(env.get_num_agents()):
+
+            # Remember and train agent
+            if Training:
+                agent.step(agent_obs[a], action_dict[a], all_rewards[a], agent_next_obs[a], done[a])
+
+            # Update the current score
+            score += all_rewards[a] / env.get_num_agents()
+
+        agent_obs = agent_next_obs.copy()
+        if done['__all__']:
+            env_done = 1
+            break
+
+    # Epsilon decay
+    eps = max(eps_end, eps_decay * eps)  # decrease epsilon
+```
+
+Running the `training_navigation.py` file trains a simple agent to navigate to any random target within the railway network. After running you should see a learning curve similiar to this one:
+
+![Learning_curve](https://i.imgur.com/yVGXpUy.png)
+
+and the agent behavior should look like this:
+
+![Single_Agent_Navigation](https://i.imgur.com/t5ULr4L.gif)
+

torch_training/Multi_Agent_Training_Intro.md

+# How to train multiple Agents on Flatland
+Quick introduction on how to train a simple DQN agent using Flatland and Pytorch. At the end of this Tutorial you should be able to train a single agent to navigate in Flatland.
+We use the `multi_agent_training.py` ([here](https://gitlab.aicrowd.com/flatland/baselines/blob/master/torch_training/multi_agent_training.py)) file to train multiple agents on the avoid conflicts task.
+
+## Actions in Flatland
+Flatland is a railway simulation. Thus the actions of an agent are strongly limited to the railway network. This means that in many cases not all actions are valid.
+The possible actions of an agent are
+
+- 0 *Do Nothing*:  If the agent is moving it continues moving, if it is stopped it stays stopped
+- 1 *Deviate Left*: This action is only valid at cells where the agent can change direction towards left. If action is chosen, the left transition and a rotation of the agent orientation to the left is executed. If the agent is stopped at any position, this action will cause it to start moving in any cell where forward or left is allowed!
+- 2 *Go Forward*: This action will start the agent when stopped. At switches this will chose the forward direction.
+- 3 *Deviate Right*: Exactly the same as deviate left but for right turns.
+- 4 *Stop*: This action causes the agent to stop, this is necessary to avoid conflicts in multi agent setups (Not needed for navigation).
+
+## Shortest path predictor
+With multiple agents alot of conlflicts will arise on the railway network. These conflicts arise because different agents want to occupie the same cells at the same time. Due to the nature of the railway network and the dynamic of the railway agents (can't turn around), the conflicts have to be detected in advance in order to avoid them. If agents are facing each other and don't have any options to deviate from their path it is called a *deadlock*.
+Therefore we introduce a simple prediction function that predicts the most likely (here shortest) path of all the agents. Furthermore, the prediction is withdrawn if an agent stopps and replaced by a prediction that the agent will stay put. The predictions allow the agents to detect possible conflicts before they happen and thus performe counter measures.
+*ATTENTION*: This is a very basic implementation of a predictor. It will not solve all the problems because it always predicts shortest paths and not alternative routes. It is up to you to come up with much more clever predictors to avod conflicts!
+
+## Tree Observation
+Flatland offers three basic observations from the beginning. We encourage you to develop your own observations that are better suited for this specific task.
+
+For the navigation training we start with the Tree Observation as agents will learn the task very quickly using this observation.
+The tree observation exploits the fact that a railway network is a graph and thus the observation is only built along allowed transitions in the graph.
+
+Here is a small example of a railway network with an agent in the top left corner. The tree observation is build by following the allowed transitions for that agent.
+
+![Small_Network](https://i.imgur.com/utqMx08.png)
+
+As we move along the allowed transitions we build up a tree where a new node is created at every cell where the agent has different possibilities (Switch), dead-end or the target is reached.
+It is important to note that the tree observation is always build according to the orientation of the agent at a given node. This means that each node always has 4 branches coming from it in the directions *Left, Forward, Right and Backward*. These are illustrated with different colors in the figure below. The tree is build form the example rail above. Nodes where there are no possibilities are filled with `-inf` and are not all shown here for simplicity. The tree however, always has the same number of nodes for a given tree depth.
+
+![Tree_Observation](https://i.imgur.com/VsUQOQz.png)
+
+### Node Information
+Each node is filled with information gathered along the path to the node. Currently each node contains 9 features:
+
+- 1: if own target lies on the explored branch the current distance from the agent in number of cells is stored.
+
+- 2: if another agents target is detected the distance in number of cells from current agent position is stored.
+
+- 3: if another agent is detected the distance in number of cells from current agent position is stored.
+
+- 4: possible conflict detected (This only works when we use a predictor and will not be important in this tutorial)
+
+
+- 5: if an not usable switch (for agent) is detected we store the distance. An unusable switch is a switch where the agent does not have any choice of path, but other agents coming from different directions might. 
+
+
+- 6: This feature stores the distance (in number of cells) to the next node (e.g. switch or target or dead-end)
+
+- 7: minimum remaining travel distance from node to the agent's target given the direction of the agent if this path is chosen
+
+
+- 8: agent in the same direction found on path to node
+    - n = number of agents present same direction (possible future use: number of other agents in the same direction in this branch)
+    - 0 = no agent present same direction
+
+- 9: agent in the opposite direction on path to node
+    - n = number of agents present other direction than myself
+    - 0 = no agent present other direction than myself
+
+For training purposes the tree is flattend into a single array.
+
+
+## Training
+### Setting up the environment
+Let us now train a simle double dueling DQN agent to detect to find its target and try to avoid conflicts on flatland. We start by importing the necessary packages from Flatland. Note that we now also import a predictor from `flatland.envs.predictions`
+
+```
+from flatland.envs.generators import complex_rail_generator
+from flatland.envs.observations import TreeObsForRailEnv
+from flatland.envs.predictions import ShortestPathPredictorForRailEnv
+from flatland.envs.rail_env import RailEnv
+from utils.observation_utils import norm_obs_clip, split_tree
+```
+
+For this simple example we want to train on randomly generated levels using the `complex_rail_generator`. The training curriculum will use different sets of parameters throughout training to enhance generalizability of the solution.
+
+```
+# Initialize a random map with a random number of agents
+x_dim = np.random.randint(8, 20)
+y_dim = np.random.randint(8, 20)
+n_agents = np.random.randint(3, 8)
+n_goals = n_agents + np.random.randint(0, 3)
+min_dist = int(0.75 * min(x_dim, y_dim))
+tree_depth = 3
+```
+
+As mentioned above, for this experiment we are going to use the tree observation and thus we load the observation builder. Also we are now using the predictor as well which is passed to the observation builder.
+
+```
+"""
+ Get an observation builder and predictor:
+ The predictor will always predict the shortest path from the current location of the agent.
+ This is used to warn for potential conflicts --> Should be enhanced to get better performance!
+"""
+predictor = ShortestPathPredictorForRailEnv()
+observation_helper = TreeObsForRailEnv(max_depth=tree_depth, predictor=predictor)
+```
+
+And pass it as an argument to the environment setup
+
+```
+env = RailEnv(width=x_dim,
+              height=y_dim,
+              rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
+                                                    max_dist=99999,
+                                                    seed=0),
+              obs_builder_object=observation_builder,
+              number_of_agents=n_agents)
+```
+
+We have no successfully set up the environment for training. To visualize it in the renderer we also initiate the renderer with.
+
+###Setting up the agent
+
+To set up a appropriate agent we need the state and action space sizes. From the discussion above about the tree observation we end up with:
+
+
+```
+num_features_per_node = env.obs_builder.observation_dim
+nr_nodes = 0
+for i in range(tree_depth + 1):
+    nr_nodes += np.power(4, i)
+state_size = num_features_per_node * nr_nodes
+action_size = 5
+```
+
+In the `multi_agent_training.py` file you will find further variable that we initiate in order to keep track of the training progress.
+Below you see an example code to train an agent. It is important to note that we reshape and normalize the tree observation provided by the environment to facilitate training.
+To do so, we use the utility functions `split_tree(tree=np.array(obs[a]), num_features_per_node=features_per_node, current_depth=0)` and `norm_obs_clip()`. Feel free to modify the normalization as you see fit.
+
+```
+# Split the observation tree into its parts and normalize the observation using the utility functions.
+    # Build agent specific local observation
+    for a in range(env.get_num_agents()):
+        rail_data, distance_data, agent_data = split_tree(tree=np.array(obs[a]),
+                                                          num_features_per_node=features_per_node,
+                                                          current_depth=0)
+        rail_data = norm_obs_clip(rail_data)
+        distance_data = norm_obs_clip(distance_data)
+        agent_data = np.clip(agent_data, -1, 1)
+        agent_obs[a] = np.concatenate((np.concatenate((rail_data, distance_data)), agent_data))
+```
+
+We now use the normalized `agent_obs` for our training loop:
+
+
+```
+# Do training over n_episodes
+    for episodes in range(1, n_episodes + 1):
+        """
+        Training Curriculum: In order to get good generalization we change the number of agents
+        and the size of the levels every 50 episodes.
+        """
+        if episodes % 50 == 0:
+            x_dim = np.random.randint(8, 20)
+            y_dim = np.random.randint(8, 20)
+            n_agents = np.random.randint(3, 8)
+            n_goals = n_agents + np.random.randint(0, 3)
+            min_dist = int(0.75 * min(x_dim, y_dim))
+            env = RailEnv(width=x_dim,
+                          height=y_dim,
+                          rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
+                                                                max_dist=99999,
+                                                                seed=0),
+                          obs_builder_object=observation_helper,
+                          number_of_agents=n_agents)
+
+            # Adjust the parameters according to the new env.
+            max_steps = int(3 * (env.height + env.width))
+            agent_obs = [None] * env.get_num_agents()
+            agent_next_obs = [None] * env.get_num_agents()
+
+        # Reset environment
+        obs, info = env.reset(True, True)
+
+        # Setup placeholder for finals observation of a single agent. This is necessary because agents terminate at
+        # different times during an episode
+        final_obs = agent_obs.copy()
+        final_obs_next = agent_next_obs.copy()
+
+        # Build agent specific observations
+        for a in range(env.get_num_agents()):
+            data, distance, agent_data = split_tree(tree=np.array(obs[a]), num_features_per_node=num_features_per_node,
+                                                    current_depth=0)
+            data = norm_obs_clip(data, fixed_radius=observation_radius)
+            distance = norm_obs_clip(distance)
+            agent_data = np.clip(agent_data, -1, 1)
+            agent_obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
+
+        score = 0
+        env_done = 0
+
+        # Run episode
+        for step in range(max_steps):
+
+            # Action
+            for a in range(env.get_num_agents()):
+                action = agent.act(agent_obs[a], eps=eps)
+                action_prob[action] += 1
+                action_dict.update({a: action})
+
+            # Environment step
+            next_obs, all_rewards, done, _ = env.step(action_dict)
+
+            # Build agent specific observations and normalize
+            for a in range(env.get_num_agents()):
+                data, distance, agent_data = split_tree(tree=np.array(next_obs[a]),
+                                                        num_features_per_node=num_features_per_node, current_depth=0)
+                data = norm_obs_clip(data, fixed_radius=observation_radius)
+                distance = norm_obs_clip(distance)
+                agent_data = np.clip(agent_data, -1, 1)
+                agent_next_obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
+
+            # Update replay buffer and train agent
+            for a in range(env.get_num_agents()):
+                if done[a]:
+                    final_obs[a] = agent_obs[a].copy()
+                    final_obs_next[a] = agent_next_obs[a].copy()
+                    final_action_dict.update({a: action_dict[a]})
+                if not done[a]:
+                    agent.step(agent_obs[a], action_dict[a], all_rewards[a], agent_next_obs[a], done[a])
+                score += all_rewards[a] / env.get_num_agents()
+
+            # Copy observation
+            agent_obs = agent_next_obs.copy()
+
+            if done['__all__']:
+                env_done = 1
+                for a in range(env.get_num_agents()):
+                    agent.step(final_obs[a], final_action_dict[a], all_rewards[a], final_obs_next[a], done[a])
+                break
+
+        # Epsilon decay
+        eps = max(eps_end, eps_decay * eps)  # decrease epsilon
+
+        # Collection information about training
+        done_window.append(env_done)
+        scores_window.append(score / max_steps)  # save most recent score
+        scores.append(np.mean(scores_window))
+        dones_list.append((np.mean(done_window)))
+```
+
+Running the `multi_agent_training.py` file trains a simple agent to navigate to any random target within the railway network. After running you should see a learning curve similiar to this one:
+
+![Learning_Curve](https://i.imgur.com/Po4j4yK.png)
+
+and the agent behavior should look like this:
+
+![Conflict_Avoidence](https://i.imgur.com/AvBHKaD.gif)

torch_training/dueling_double_dqn.py

@@ -8,51 +8,41 @@ import torch
 import torch.nn.functional as F
 import torch.optim as optim
 
-from torch_training.model import QNetwork, QNetwork2
+from torch_training.model import QNetwork
 
 BUFFER_SIZE = int(1e5)  # replay buffer size
 BATCH_SIZE = 512  # minibatch size
 GAMMA = 0.99  # discount factor 0.99
 TAU = 1e-3  # for soft update of target parameters
-LR = 0.5e-4  # learning rate 5
+LR = 0.5e-4  # learning rate 0.5e-4 works
 UPDATE_EVERY = 10  # how often to update the network
-double_dqn = True  # If using double dqn algorithm
-input_channels = 5  # Number of Input channels
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-device = torch.device("cpu")
 print(device)
 
 
 class Agent:
     """Interacts with and learns from the environment."""
 
-    def __init__(self, state_size, action_size, net_type, seed, double_dqn=True, input_channels=5):
+    def __init__(self, state_size, action_size, double_dqn=True):
         """Initialize an Agent object.
 
         Params
         ======
             state_size (int): dimension of each state
             action_size (int): dimension of each action
-            seed (int): random seed
         """
         self.state_size = state_size
         self.action_size = action_size
-        self.seed = random.seed(seed)
-        self.version = net_type
         self.double_dqn = double_dqn
         # Q-Network
-        if self.version == "Conv":
-            self.qnetwork_local = QNetwork2(state_size, action_size, seed, input_channels).to(device)
-            self.qnetwork_target = copy.deepcopy(self.qnetwork_local)
-        else:
-            self.qnetwork_local = QNetwork(state_size, action_size, seed).to(device)
-            self.qnetwork_target = copy.deepcopy(self.qnetwork_local)
+        self.qnetwork_local = QNetwork(state_size, action_size).to(device)
+        self.qnetwork_target = copy.deepcopy(self.qnetwork_local)
 
         self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
 
         # Replay memory
-        self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
+        self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE)
         # Initialize time step (for updating every UPDATE_EVERY steps)
         self.t_step = 0
 
@@ -152,7 +142,7 @@ class Agent:
 class ReplayBuffer:
     """Fixed-size buffer to store experience tuples."""
 
-    def __init__(self, action_size, buffer_size, batch_size, seed):
+    def __init__(self, action_size, buffer_size, batch_size):
         """Initialize a ReplayBuffer object.
 
         Params
@@ -160,13 +150,11 @@ class ReplayBuffer:
             action_size (int): dimension of each action
             buffer_size (int): maximum size of buffer
             batch_size (int): size of each training batch
-            seed (int): random seed
         """
         self.action_size = action_size
         self.memory = deque(maxlen=buffer_size)
         self.batch_size = batch_size
         self.experience = namedtuple("Experience", field_names=["state", "action", "reward", "next_state", "done"])
-        self.seed = random.seed(seed)
 
     def add(self, state, action, reward, next_state, done):
         """Add a new experience to memory."""
@@ -188,7 +176,7 @@ class ReplayBuffer:
         dones = torch.from_numpy(self.__v_stack_impr([e.done for e in experiences if e is not None]).astype(np.uint8)) \
             .float().to(device)
 
-        return (states, actions, rewards, next_states, dones)
+        return states, actions, rewards, next_states, dones
 
     def __len__(self):
         """Return the current size of internal memory."""

torch_training/model.py

@@ -3,7 +3,7 @@ import torch.nn.functional as F
 
 
 class QNetwork(nn.Module):
-    def __init__(self, state_size, action_size, seed, hidsize1=128, hidsize2=128):
+    def __init__(self, state_size, action_size, hidsize1=128, hidsize2=128):
         super(QNetwork, self).__init__()
 
         self.fc1_val = nn.Linear(state_size, hidsize1)
@@ -24,38 +24,3 @@ class QNetwork(nn.Module):
         adv = F.relu(self.fc2_adv(adv))
         adv = self.fc3_adv(adv)
         return val + adv - adv.mean()
-
-
-class QNetwork2(nn.Module):
-    def __init__(self, state_size, action_size, seed, input_channels, hidsize1=128, hidsize2=64):
-        super(QNetwork2, self).__init__()
-        self.conv1 = nn.Conv2d(input_channels, 16, kernel_size=3, stride=1)
-        self.bn1 = nn.BatchNorm2d(16)
-        self.conv2 = nn.Conv2d(16, 32, kernel_size=5, stride=3)
-        self.bn2 = nn.BatchNorm2d(32)
-        self.conv3 = nn.Conv2d(32, 64, kernel_size=5, stride=3)
-        self.bn3 = nn.BatchNorm2d(64)
-
-        self.fc1_val = nn.Linear(6400, hidsize1)
-        self.fc2_val = nn.Linear(hidsize1, hidsize2)
-        self.fc3_val = nn.Linear(hidsize2, 1)
-
-        self.fc1_adv = nn.Linear(6400, hidsize1)
-        self.fc2_adv = nn.Linear(hidsize1, hidsize2)
-        self.fc3_adv = nn.Linear(hidsize2, action_size)
-
-    def forward(self, x):
-        x = F.relu(self.conv1(x))
-        x = F.relu(self.conv2(x))
-        x = F.relu(self.conv3(x))
-
-        # value function approximation
-        val = F.relu(self.fc1_val(x.view(x.size(0), -1)))
-        val = F.relu(self.fc2_val(val))
-        val = self.fc3_val(val)
-
-        # advantage calculation
-        adv = F.relu(self.fc1_adv(x.view(x.size(0), -1)))
-        adv = F.relu(self.fc2_adv(adv))
-        adv = self.fc3_adv(adv)
-        return val + adv - adv.mean()

torch_training/multi_agent_inference.py

+import random
+from collections import deque
+
+import numpy as np
+import torch
+from flatland.envs.malfunction_generators import malfunction_from_params, MalfunctionParameters
+from flatland.envs.observations import TreeObsForRailEnv
+from flatland.envs.predictions import ShortestPathPredictorForRailEnv
+from flatland.envs.rail_env import RailEnv
+from flatland.envs.rail_generators import sparse_rail_generator
+from flatland.envs.schedule_generators import sparse_schedule_generator
+from flatland.utils.rendertools import RenderTool
+from importlib_resources import path
+
+import torch_training.Nets
+from torch_training.dueling_double_dqn import Agent
+from utils.observation_utils import normalize_observation
+
+random.seed(1)
+np.random.seed(1)
+"""
+file_name = "./railway/complex_scene.pkl"
+env = RailEnv(width=10,
+              height=20,
+              rail_generator=rail_from_file(file_name),
+              obs_builder_object=TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv()))
+x_dim = env.width
+y_dim = env.height
+"""
+
+# Parameters for the Environment
+x_dim = 25
+y_dim = 25
+n_agents = 10
+
+# We are training an Agent using the Tree Observation with depth 2
+observation_builder = TreeObsForRailEnv(max_depth=2)
+
+# Use a the malfunction generator to break agents from time to time
+stochastic_data = MalfunctionParameters(malfunction_rate=1./10000,  # Rate of malfunction occurence
+                                        min_duration=15,  # Minimal duration of malfunction
+                                        max_duration=50  # Max duration of malfunction
+                                        )
+
+
+
+# Custom observation builder
+TreeObservation = TreeObsForRailEnv(max_depth=2, predictor=ShortestPathPredictorForRailEnv(30))
+
+# Different agent types (trains) with different speeds.
+speed_ration_map = {1.: 0.25,  # Fast passenger train
+                    1. / 2.: 0.25,  # Fast freight train
+                    1. / 3.: 0.25,  # Slow commuter train
+                    1. / 4.: 0.25}  # Slow freight train
+
+env = RailEnv(width=x_dim,
+              height=y_dim,
+              rail_generator=sparse_rail_generator(max_num_cities=3,
+                                                   # Number of cities in map (where train stations are)
+                                                   seed=1,  # Random seed
+                                                   grid_mode=False,
+                                                   max_rails_between_cities=2,
+                                                   max_rails_in_city=2),
+              schedule_generator=sparse_schedule_generator(speed_ration_map),
+              number_of_agents=n_agents,
+              malfunction_generator_and_process_data=malfunction_from_params(stochastic_data),
+              obs_builder_object=TreeObservation)
+env.reset(True, True)
+
+observation_helper = TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv())
+env_renderer = RenderTool(env, gl="PILSVG", )
+num_features_per_node = env.obs_builder.observation_dim
+
+tree_depth = 2
+nr_nodes = 0
+for i in range(tree_depth + 1):
+    nr_nodes += np.power(4, i)
+state_size = num_features_per_node * nr_nodes
+action_size = 5
+
+# We set the number of episodes we would like to train on
+if 'n_trials' not in locals():
+    n_trials = 60000
+max_steps = int(4 * 2 * (20 + env.height + env.width))
+eps = 1.
+eps_end = 0.005
+eps_decay = 0.9995
+action_dict = dict()
+final_action_dict = dict()
+scores_window = deque(maxlen=100)
+done_window = deque(maxlen=100)
+scores = []
+dones_list = []
+action_prob = [0] * action_size
+agent_obs = [None] * env.get_num_agents()
+agent_next_obs = [None] * env.get_num_agents()
+agent = Agent(state_size, action_size)
+with path(torch_training.Nets, "navigator_checkpoint1200.pth") as file_in:
+    agent.qnetwork_local.load_state_dict(torch.load(file_in))
+
+record_images = False
+frame_step = 0
+
+for trials in range(1, n_trials + 1):
+
+    # Reset environment
+    obs, info = env.reset(True, True)
+    env_renderer.reset()
+    # Build agent specific observations
+    for a in range(env.get_num_agents()):
+        agent_obs[a] = agent_obs[a] = normalize_observation(obs[a], tree_depth, observation_radius=10)
+    # Reset score and done
+    score = 0
+    env_done = 0
+
+    # Run episode
+    for step in range(max_steps):
+
+        # Action
+        for a in range(env.get_num_agents()):
+            if info['action_required'][a]:
+                action = agent.act(agent_obs[a], eps=0.)
+            else:
+                action = 0
+
+            action_prob[action] += 1
+            action_dict.update({a: action})
+        # Environment step
+        obs, all_rewards, done, _ = env.step(action_dict)
+        env_renderer.render_env(show=True, show_predictions=True, show_observations=False)
+        # Build agent specific observations and normalize
+        for a in range(env.get_num_agents()):
+            if obs[a]:
+                agent_obs[a] = normalize_observation(obs[a], tree_depth, observation_radius=10)
+
+
+        if done['__all__']:
+            break
+

torch_training/multi_agent_training.py

+import getopt
 import random
+import sys
 from collections import deque
+# make sure the root path is in system path
+from pathlib import Path
+
+from flatland.envs.malfunction_generators import malfunction_from_params, MalfunctionParameters
+
+base_dir = Path(__file__).resolve().parent.parent
+sys.path.append(str(base_dir))
 
 import matplotlib.pyplot as plt
 import numpy as np
 import torch
-from dueling_double_dqn import Agent
-from flatland.envs.generators import complex_rail_generator
-from flatland.envs.observations import TreeObsForRailEnv
-from flatland.envs.predictions import ShortestPathPredictorForRailEnv
+from torch_training.dueling_double_dqn import Agent
+
 from flatland.envs.rail_env import RailEnv
+from flatland.envs.rail_generators import sparse_rail_generator
+from flatland.envs.schedule_generators import sparse_schedule_generator
 from flatland.utils.rendertools import RenderTool
-
-from utils.observation_utils import norm_obs_clip, split_tree
-
-random.seed(1)
-np.random.seed(1)
-
-"""
-env = RailEnv(width=10,
-              height=20, obs_builder_object=TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv()))
-env.load("./railway/complex_scene.pkl")
-file_load = True
-"""
-
-x_dim = np.random.randint(8, 20)
-y_dim = np.random.randint(8, 20)
-n_agents = np.random.randint(3, 8)
-n_goals = n_agents + np.random.randint(0, 3)
-min_dist = int(0.75 * min(x_dim, y_dim))
-env = RailEnv(width=x_dim,
-              height=y_dim,
-              rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
-                                                    max_dist=99999,
-                                                    seed=0),
-              obs_builder_object=TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv()),
-              number_of_agents=n_agents)
-env.reset(True, True)
-file_load = False
-"""
-
-"""
-observation_helper = TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv())
-env_renderer = RenderTool(env, gl="PILSVG", )
-handle = env.get_agent_handles()
-features_per_node = 9
-state_size = features_per_node * 85 * 2
-action_size = 5
-n_trials = 30000
-max_steps = int(3 * (env.height + env.width))
-eps = 1.
-eps_end = 0.005
-eps_decay = 0.9995
-action_dict = dict()
-final_action_dict = dict()
-scores_window = deque(maxlen=100)
-done_window = deque(maxlen=100)
-time_obs = deque(maxlen=2)
-scores = []
-dones_list = []
-action_prob = [0] * action_size
-agent_obs = [None] * env.get_num_agents()
-agent_next_obs = [None] * env.get_num_agents()
-agent = Agent(state_size, action_size, "FC", 0)
-agent.qnetwork_local.load_state_dict(torch.load('./Nets/avoid_checkpoint30000.pth'))
-
-demo = True
-record_images = False
-
-for trials in range(1, n_trials + 1):
-
-    if trials % 50 == 0 and not demo:
-        x_dim = np.random.randint(8, 20)
-        y_dim = np.random.randint(8, 20)
-        n_agents = np.random.randint(3, 8)
-        n_goals = n_agents + np.random.randint(0, 3)
-        min_dist = int(0.75 * min(x_dim, y_dim))
-        env = RailEnv(width=x_dim,
-                      height=y_dim,
-                      rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
-                                                            max_dist=99999,
-                                                            seed=0),
-                      obs_builder_object=TreeObsForRailEnv(max_depth=3, predictor=ShortestPathPredictorForRailEnv()),
-                      number_of_agents=n_agents)
-        env.reset(True, True)
-        max_steps = int(3 * (env.height + env.width))
-        agent_obs = [None] * env.get_num_agents()
-        agent_next_obs = [None] * env.get_num_agents()
-    # Reset environment
-    if file_load:
-        obs = env.reset(False, False)
-    else:
-        obs = env.reset(True, True)
-    if demo:
-        env_renderer.set_new_rail()
-    obs_original = obs.copy()
-    final_obs = obs.copy()
-    final_obs_next = obs.copy()
-    for a in range(env.get_num_agents()):
-        data, distance, agent_data = split_tree(tree=np.array(obs[a]), num_features_per_node=features_per_node,
-                                                current_depth=0)
-        data = norm_obs_clip(data)
-        distance = norm_obs_clip(distance)
-        agent_data = np.clip(agent_data, -1, 1)
-        obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
-        agent_data = env.agents[a]
-        speed = 1  # np.random.randint(1,5)
-        agent_data.speed_data['speed'] = 1. / speed
-
-    for i in range(2):
-        time_obs.append(obs)
-    # env.obs_builder.util_print_obs_subtree(tree=obs[0], num_elements_per_node=5)
-    for a in range(env.get_num_agents()):
-        agent_obs[a] = np.concatenate((time_obs[0][a], time_obs[1][a]))
-
-    score = 0
-    env_done = 0
-    # Run episode
-    for step in range(max_steps):
-        if demo:
-            env_renderer.renderEnv(show=True, show_observations=True)
-            # observation_helper.util_print_obs_subtree(obs_original[0])
-            if record_images:
-                env_renderer.gl.saveImage("./Images/flatland_frame_{:04d}.bmp".format(step))
-        # print(step)
-        # Action
-        for a in range(env.get_num_agents()):
-            if demo:
-                eps = 0
-            # action = agent.act(np.array(obs[a]), eps=eps)
-            action = agent.act(agent_obs[a], eps=eps)
-            action_prob[action] += 1
-            action_dict.update({a: action})
-        # Environment step
-
-        next_obs, all_rewards, done, _ = env.step(action_dict)
-        # print(all_rewards,action)
-        obs_original = next_obs.copy()
-        for a in range(env.get_num_agents()):
-            data, distance, agent_data = split_tree(tree=np.array(next_obs[a]), num_features_per_node=features_per_node,
-                                                    current_depth=0)
-            data = norm_obs_clip(data)
-            distance = norm_obs_clip(distance)
-            agent_data = np.clip(agent_data, -1, 1)
-            next_obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
-        time_obs.append(next_obs)
-
-        # Update replay buffer and train agent
+from utils.observation_utils import normalize_observation
+from flatland.envs.observations import TreeObsForRailEnv
+from flatland.envs.predictions import ShortestPathPredictorForRailEnv
+from flatland.envs.agent_utils import RailAgentStatus
+
+def main(argv):
+    try:
+        opts, args = getopt.getopt(argv, "n:", ["n_trials="])
+    except getopt.GetoptError:
+        print('training_navigation.py -n <n_trials>')
+        sys.exit(2)
+    for opt, arg in opts:
+        if opt in ('-n', '--n_trials'):
+            n_trials = int(arg)
+
+    random.seed(1)
+    np.random.seed(1)
+
+    # Parameters for the Environment
+    x_dim = 35
+    y_dim = 35
+    n_agents = 10
+
+
+    # Use a the malfunction generator to break agents from time to time
+    stochastic_data = MalfunctionParameters(malfunction_rate=1./10000,  # Rate of malfunction occurence
+                                            min_duration=15,  # Minimal duration of malfunction
+                                            max_duration=50  # Max duration of malfunction
+                                            )
+
+
+    # Custom observation builder
+    TreeObservation = TreeObsForRailEnv(max_depth=2, predictor=ShortestPathPredictorForRailEnv(30))
+
+    # Different agent types (trains) with different speeds.
+    speed_ration_map = {1.: 0.25,  # Fast passenger train
+                        1. / 2.: 0.25,  # Fast freight train
+                        1. / 3.: 0.25,  # Slow commuter train
+                        1. / 4.: 0.25}  # Slow freight train
+
+    env = RailEnv(width=x_dim,
+                  height=y_dim,
+                  rail_generator=sparse_rail_generator(max_num_cities=3,
+                                                       # Number of cities in map (where train stations are)
+                                                       seed=1,  # Random seed
+                                                       grid_mode=False,
+                                                       max_rails_between_cities=2,
+                                                       max_rails_in_city=3),
+                  schedule_generator=sparse_schedule_generator(speed_ration_map),
+                  number_of_agents=n_agents,
+                  malfunction_generator_and_process_data=malfunction_from_params(stochastic_data),
+                  obs_builder_object=TreeObservation)
+
+    # Reset env
+    env.reset(True,True)
+    # After training we want to render the results so we also load a renderer
+    env_renderer = RenderTool(env, gl="PILSVG", )
+    # Given the depth of the tree observation and the number of features per node we get the following state_size
+    num_features_per_node = env.obs_builder.observation_dim
+    tree_depth = 2
+    nr_nodes = 0
+    for i in range(tree_depth + 1):
+        nr_nodes += np.power(4, i)
+    state_size = num_features_per_node * nr_nodes
+
+    # The action space of flatland is 5 discrete actions
+    action_size = 5
+
+    # We set the number of episodes we would like to train on
+    if 'n_trials' not in locals():
+        n_trials = 15000
+
+    # And the max number of steps we want to take per episode
+    max_steps = int(4 * 2 * (20 + env.height + env.width))
+
+    # Define training parameters
+    eps = 1.
+    eps_end = 0.005
+    eps_decay = 0.998
+
+    # And some variables to keep track of the progress
+    action_dict = dict()
+    final_action_dict = dict()
+    scores_window = deque(maxlen=100)
+    done_window = deque(maxlen=100)
+    scores = []
+    dones_list = []
+    action_prob = [0] * action_size
+    agent_obs = [None] * env.get_num_agents()
+    agent_next_obs = [None] * env.get_num_agents()
+    agent_obs_buffer = [None] * env.get_num_agents()
+    agent_action_buffer = [2] * env.get_num_agents()
+    cummulated_reward = np.zeros(env.get_num_agents())
+    update_values = [False] * env.get_num_agents()
+    # Now we load a Double dueling DQN agent
+    agent = Agent(state_size, action_size)
+
+    for trials in range(1, n_trials + 1):
+
+        # Reset environment
+        obs, info = env.reset(True, True)
+        env_renderer.reset()
+        # Build agent specific observations
         for a in range(env.get_num_agents()):
-            agent_next_obs[a] = np.concatenate((time_obs[0][a], time_obs[1][a]))
-            if done[a]:
-                final_obs[a] = agent_obs[a].copy()
-                final_obs_next[a] = agent_next_obs[a].copy()
-                final_action_dict.update({a: action_dict[a]})
-            if not demo and not done[a]:
-                agent.step(agent_obs[a], action_dict[a], all_rewards[a], agent_next_obs[a], done[a])
-            score += all_rewards[a] / env.get_num_agents()
-
-        agent_obs = agent_next_obs.copy()
-        if done['__all__']:
-            env_done = 1
+            if obs[a]:
+                agent_obs[a] = normalize_observation(obs[a], tree_depth, observation_radius=10)
+                agent_obs_buffer[a] = agent_obs[a].copy()
+
+        # Reset score and done
+        score = 0
+        env_done = 0
+
+        # Run episode
+        while True:
+            # Action
+            for a in range(env.get_num_agents()):
+                if info['action_required'][a]:
+                    # If an action is require, we want to store the obs a that step as well as the action
+                    update_values[a] = True
+                    action = agent.act(agent_obs[a], eps=eps)
+                    action_prob[action] += 1
+                else:
+                    update_values[a] = False
+                    action = 0
+                action_dict.update({a: action})
+
+            # Environment step
+            next_obs, all_rewards, done, info = env.step(action_dict)
+            # Update replay buffer and train agent
             for a in range(env.get_num_agents()):
-                agent.step(final_obs[a], final_action_dict[a], all_rewards[a], final_obs_next[a], done[a])
-            break
-    # Epsilon decay
-    eps = max(eps_end, eps_decay * eps)  # decrease epsilon
-
-    done_window.append(env_done)
-    scores_window.append(score / max_steps)  # save most recent score
-    scores.append(np.mean(scores_window))
-    dones_list.append((np.mean(done_window)))
-
-    print(
-        '\rTraining {} Agents on ({},{}).\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
-            env.get_num_agents(), x_dim, y_dim,
-            trials,
-            np.mean(scores_window),
-            100 * np.mean(done_window),
-            eps, action_prob / np.sum(action_prob)), end=" ")
-
-    if trials % 100 == 0:
+                # Only update the values when we are done or when an action was taken and thus relevant information is present
+                if update_values[a] or done[a]:
+                    agent.step(agent_obs_buffer[a], agent_action_buffer[a], all_rewards[a],
+                               agent_obs[a], done[a])
+                    cummulated_reward[a] = 0.
+
+                    agent_obs_buffer[a] = agent_obs[a].copy()
+                    agent_action_buffer[a] = action_dict[a]
+                if next_obs[a]:
+                    agent_obs[a] = normalize_observation(next_obs[a], tree_depth, observation_radius=10)
+
+                score += all_rewards[a] / env.get_num_agents()
+
+            # Copy observation
+            if done['__all__']:
+                env_done = 1
+                break
+
+        # Epsilon decay
+        eps = max(eps_end, eps_decay * eps)  # decrease epsilon
+
+        # Collection information about training
+        tasks_finished = 0
+        for current_agent in env.agents:
+            if current_agent.status == RailAgentStatus.DONE_REMOVED:
+                tasks_finished += 1
+        done_window.append(tasks_finished / max(1, env.get_num_agents()))
+        scores_window.append(score / max_steps)  # save most recent score
+        scores.append(np.mean(scores_window))
+        dones_list.append((np.mean(done_window)))
+
         print(
-            '\rTraining {} Agents.\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
-                env.get_num_agents(),
+            '\rTraining {} Agents on ({},{}).\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+                env.get_num_agents(), x_dim, y_dim,
                 trials,
                 np.mean(scores_window),
                 100 * np.mean(done_window),
-                eps,
-                action_prob / np.sum(action_prob)))
-        torch.save(agent.qnetwork_local.state_dict(),
-                   './Nets/avoid_checkpoint' + str(trials) + '.pth')
-        action_prob = [1] * action_size
-plt.plot(scores)
-plt.show()
+                eps, action_prob / np.sum(action_prob)), end=" ")
+
+        if trials % 100 == 0:
+            print(
+                '\rTraining {} Agents on ({},{}).\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+                    env.get_num_agents(), x_dim, y_dim,
+                    trials,
+                    np.mean(scores_window),
+                    100 * np.mean(done_window),
+                    eps, action_prob / np.sum(action_prob)))
+            torch.save(agent.qnetwork_local.state_dict(),
+                       './Nets/navigator_checkpoint' + str(trials) + '.pth')
+            action_prob = [1] * action_size
+
+    # Plot overall training progress at the end
+    plt.plot(scores)
+    plt.show()
+
+
+if __name__ == '__main__':
+    main(sys.argv[1:])

torch_training/multi_agent_two_time_step_training.py

+# Import packages for plotting and system
+import getopt
+import random
+import sys
+from collections import deque
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from flatland.envs.observations import TreeObsForRailEnv
+from flatland.envs.predictions import ShortestPathPredictorForRailEnv
+from flatland.envs.rail_env import RailEnv
+from flatland.envs.rail_generators import complex_rail_generator
+# Import Flatland/ Observations and Predictors
+from flatland.envs.schedule_generators import complex_schedule_generator
+from importlib_resources import path
+
+# Import Torch and utility functions to normalize observation
+import torch_training.Nets
+from torch_training.dueling_double_dqn import Agent
+from utils.observation_utils import norm_obs_clip, split_tree_into_feature_groups
+
+
+def main(argv):
+    try:
+        opts, args = getopt.getopt(argv, "n:", ["n_episodes="])
+    except getopt.GetoptError:
+        print('training_navigation.py -n <n_episodes>')
+        sys.exit(2)
+    for opt, arg in opts:
+        if opt in ('-n', '--n_episodes'):
+            n_episodes = int(arg)
+
+    ## Initialize the random
+    random.seed(1)
+    np.random.seed(1)
+
+    # Initialize a random map with a random number of agents
+    x_dim = np.random.randint(8, 20)
+    y_dim = np.random.randint(8, 20)
+    n_agents = np.random.randint(3, 8)
+    n_goals = n_agents + np.random.randint(0, 3)
+    min_dist = int(0.75 * min(x_dim, y_dim))
+    tree_depth = 2
+    print("main2")
+    demo = False
+
+    # Get an observation builder and predictor
+    observation_helper = TreeObsForRailEnv(max_depth=tree_depth, predictor=ShortestPathPredictorForRailEnv())
+
+    env = RailEnv(width=x_dim,
+                  height=y_dim,
+                  rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
+                                                        max_dist=99999,
+                                                        seed=0),
+                  schedule_generator=complex_schedule_generator(),
+                  obs_builder_object=observation_helper,
+                  number_of_agents=n_agents)
+    env.reset(True, True)
+
+    handle = env.get_agent_handles()
+    features_per_node = env.obs_builder.observation_dim
+    nr_nodes = 0
+    for i in range(tree_depth + 1):
+        nr_nodes += np.power(4, i)
+    state_size = 2 * features_per_node * nr_nodes  # We will use two time steps per observation --> 2x state_size
+    action_size = 5
+
+    # We set the number of episodes we would like to train on
+    if 'n_episodes' not in locals():
+        n_episodes = 60000
+
+    # Set max number of steps per episode as well as other training relevant parameter
+    max_steps = int(3 * (env.height + env.width))
+    eps = 1.
+    eps_end = 0.005
+    eps_decay = 0.9995
+    action_dict = dict()
+    final_action_dict = dict()
+    scores_window = deque(maxlen=100)
+    done_window = deque(maxlen=100)
+    time_obs = deque(maxlen=2)
+    scores = []
+    dones_list = []
+    action_prob = [0] * action_size
+    agent_obs = [None] * env.get_num_agents()
+    agent_next_obs = [None] * env.get_num_agents()
+    # Initialize the agent
+    agent = Agent(state_size, action_size)
+
+    # Here you can pre-load an agent
+    if False:
+        with path(torch_training.Nets, "avoid_checkpoint500.pth") as file_in:
+            agent.qnetwork_local.load_state_dict(torch.load(file_in))
+
+    # Do training over n_episodes
+    for episodes in range(1, n_episodes + 1):
+        """
+        Training Curriculum: In order to get good generalization we change the number of agents
+        and the size of the levels every 50 episodes.
+        """
+        if episodes % 50 == 0:
+            x_dim = np.random.randint(8, 20)
+            y_dim = np.random.randint(8, 20)
+            n_agents = np.random.randint(3, 8)
+            n_goals = n_agents + np.random.randint(0, 3)
+            min_dist = int(0.75 * min(x_dim, y_dim))
+            env = RailEnv(width=x_dim,
+                          height=y_dim,
+                          rail_generator=complex_rail_generator(nr_start_goal=n_goals, nr_extra=5, min_dist=min_dist,
+                                                                max_dist=99999,
+                                                                seed=0),
+                          schedule_generator=complex_schedule_generator(),
+                          obs_builder_object=TreeObsForRailEnv(max_depth=3,
+                                                               predictor=ShortestPathPredictorForRailEnv()),
+                          number_of_agents=n_agents)
+
+            # Adjust the parameters according to the new env.
+            max_steps = int(3 * (env.height + env.width))
+            agent_obs = [None] * env.get_num_agents()
+            agent_next_obs = [None] * env.get_num_agents()
+
+        # Reset environment
+        obs, info = env.reset(True, True)
+
+        # Setup placeholder for finals observation of a single agent. This is necessary because agents terminate at
+        # different times during an episode
+        final_obs = agent_obs.copy()
+        final_obs_next = agent_next_obs.copy()
+
+        # Build agent specific observations
+        for a in range(env.get_num_agents()):
+            data, distance, agent_data = split_tree_into_feature_groups(obs[a], tree_depth)
+            data = norm_obs_clip(data)
+            distance = norm_obs_clip(distance)
+            agent_data = np.clip(agent_data, -1, 1)
+            obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
+
+        # Accumulate two time steps of observation (Here just twice the first state)
+        for i in range(2):
+            time_obs.append(obs)
+
+        # Build the agent specific double ti
+        for a in range(env.get_num_agents()):
+            agent_obs[a] = np.concatenate((time_obs[0][a], time_obs[1][a]))
+
+        score = 0
+        env_done = 0
+        # Run episode
+        for step in range(max_steps):
+
+            # Action
+            for a in range(env.get_num_agents()):
+                if demo:
+                    eps = 0
+                # action = agent.act(np.array(obs[a]), eps=eps)
+                action = agent.act(agent_obs[a], eps=eps)
+                action_prob[action] += 1
+                action_dict.update({a: action})
+            # Environment step
+
+            next_obs, all_rewards, done, _ = env.step(action_dict)
+            for a in range(env.get_num_agents()):
+                data, distance, agent_data = split_tree_into_feature_groups(next_obs[a], tree_depth)
+                data = norm_obs_clip(data)
+                distance = norm_obs_clip(distance)
+                agent_data = np.clip(agent_data, -1, 1)
+                next_obs[a] = np.concatenate((np.concatenate((data, distance)), agent_data))
+            time_obs.append(next_obs)
+
+            # Update replay buffer and train agent
+            for a in range(env.get_num_agents()):
+                agent_next_obs[a] = np.concatenate((time_obs[0][a], time_obs[1][a]))
+                if done[a]:
+                    final_obs[a] = agent_obs[a].copy()
+                    final_obs_next[a] = agent_next_obs[a].copy()
+                    final_action_dict.update({a: action_dict[a]})
+                if not demo and not done[a]:
+                    agent.step(agent_obs[a], action_dict[a], all_rewards[a], agent_next_obs[a], done[a])
+                score += all_rewards[a] / env.get_num_agents()
+
+            agent_obs = agent_next_obs.copy()
+            if done['__all__']:
+                env_done = 1
+                for a in range(env.get_num_agents()):
+                    agent.step(final_obs[a], final_action_dict[a], all_rewards[a], final_obs_next[a], done[a])
+                break
+        # Epsilon decay
+        eps = max(eps_end, eps_decay * eps)  # decrease epsilon
+
+        done_window.append(env_done)
+        scores_window.append(score / max_steps)  # save most recent score
+        scores.append(np.mean(scores_window))
+        dones_list.append((np.mean(done_window)))
+
+        print(
+            '\rTraining {} Agents on ({},{}).\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+                env.get_num_agents(), x_dim, y_dim,
+                episodes,
+                np.mean(scores_window),
+                100 * np.mean(done_window),
+                eps, action_prob / np.sum(action_prob)), end=" ")
+
+        if episodes % 100 == 0:
+            print(
+                '\rTraining {} Agents.\t Episode {}\t Average Score: {:.3f}\tDones: {:.2f}%\tEpsilon: {:.2f} \t Action Probabilities: \t {}'.format(
+                    env.get_num_agents(),
+                    episodes,
+                    np.mean(scores_window),
+                    100 * np.mean(done_window),
+                    eps,
+                    action_prob / np.sum(action_prob)))
+            torch.save(agent.qnetwork_local.state_dict(),
+                       './Nets/avoid_checkpoint' + str(episodes) + '.pth')
+            action_prob = [1] * action_size
+    plt.plot(scores)
+    plt.show()
+
+
+if __name__ == '__main__':
+    main(sys.argv[1:])