The Grid Sensor combines the generality of data extraction from Raycasts with the image processing power of Convolutional Neural Networks. The Grid Sensor can be used to collect data in the general form of a "Width x Height x Channel" matrix which can be used for training Reinforcement Learning agents or for data analysis.
The Grid Sensor is an alternative method for collecting observations which combines the generality of data extraction from Raycasts with the image processing power of Convolutional Neural Networks. The Grid Sensor can be used to collect data in the general form of a "Width x Height x Channel" matrix which can be used for training agent policies or for data analysis.
In MLAgents there are 2 main sensors for observing information that is "physically" around the agent.
In ML-Agents there are two main sensors for observing information that is "physically" around the agent.
This is simple to implement and provides enough information for most simple games. When few are used, they are computationally fast. However, there are multiple limiting factors:
* The rays need to be at the same height as the things the agent should observe
* Objects can remain hidden by line of sight and if the knowledge of those objects is crucial to the success of the agent, then this limitation must be compensated for by the agents networks capacity (i.e., need a bigger brain with memory)
Raycasts are simple to implement and provides enough information for most simple games. When few are used, they are also computationally lightweight. However, there are multiple limiting factors:
* The rays need to be at the same height as the things the agent should observe.
* Objects can remain hidden by line of sight and if the knowledge of those objects is crucial to the success of the agent, then this limitation must be compensated for by the agents networks capacity (i.e., need a bigger brain with memory).
* Typically the length of the raycasts is limited because the agent need not know about objects that are at the other side of the level. Combined with few raycasts for computational efficiency, this means that an agent may not observe objects that fall between these rays and the issue becomes worse as the objects reduce in size.
* Typically, the length of the raycasts is limited because the agent need not know about objects that are at the other side of the level. Combined with few raycasts for computational efficiency, this means that an agent may not observe objects that fall between these rays and the issue becomes worse as the objects reduce in size.
The Camera provides the agent with either a grayscale or an RGB image of the game environment. It goes without saying that there non-linear relationships between nearby pixels in an image. It is this intuition that helps form the basis of Convolutional Neural Networks (CNNs) and established the literature of designing networks that take advantage of these relationships between pixels. Following this established literature of CNNs on image based data, the MLAgent's Camera Sensor provides a means by which the agent can include high dimensional inputs (images) into its observation stream.
The Camera provides the agent with either a grayscale or an RGB image of the game environment. In many cases, what we want to extract from a set of pixels is invariant to the location of those pixels in the image. It is this intuition that helps form the basis of Convolutional Neural Networks (CNNs) and established the literature of designing networks that take advantage of these relationships between pixels. Following this established literature of CNNs on image based data, the ML-Agent's Camera Sensor provides a means by which the agent can include high dimensional inputs (images) into its observation stream.
* It requires render the scene and thus is computationally slower than alternatives that do not use rendering
* It has yet been shown that the Camera Sensor can be used on a headless machine which means it is not yet possible (if at all) to train an agent on a headless infrastructure.
* It requires rendering the scene and thus is computationally slower than alternatives that do not use rendering.
* The RGB of the camera only provides a maximum of 3 channels to the agent.
* The RGB of the camera only provides a maximum of three channels to the agent.
An image can be thought of as a matrix of a predefined width (W) and a height (H) and each pixel can be thought of as simply an array of length 3 (in the case of RGB), `[Red, Green, Blue]` holding the different channel information of the color (channel) intensities at that pixel location. Thus an image is just a 3 dimensional matrix of size WxHx3. A Grid Observation can be thought of as a generalization of this setup where in place of a pixel there is a "cell" which is an array of length N representing different channel intensities at that cell position. From a Convolutional Neural Network point of view, the introduction of multiple channels in an "image" isn't a new concept. One such example is using an RGB-Depth image which is used in several robotics applications. The distinction of Grid Observations is what the data within the channels represents. Instead of limiting the channels to color intensities, the channels within a cell of a Grid Observation generalize to any data that can be represented by a single number (float or int).
Before jumping into the details of the Grid Sensor, an important thing to note is the agent performance and qualitatively different behavior over raycasts. Unity MLAgent's comes with a suite of example environments. One in particular, the [Food Collector](https://github.com/Unity-Technologies/ml-agents/tree/release_12_docs/docs/Learning-Environment-Examples.md#food-collector), has been the focus of the Grid Sensor development.
## Overview
The Food Collector environment can be described as:
* Set-up: A multi-agent environment where agents compete to collect food.
* Goal: The agents must learn to collect as many green food spheres as possible while avoiding red spheres.
* Agents: The environment contains 5 agents with same Behavior Parameters.
There are three main phases to the observation process of the Grid Sensor:
When applying the Grid Sensor to this environment, in place of the Raycast Vector Sensor or the Camera Sensor, a Mean Reward of 40-50 is observed. This performance is on par with what is seen by agents trained with RayCasts but the side-by-side comparison of trained agents, shows a qualitative difference in behavior. A deeper study and interpretation of the qualitative differences between agents trained with Raycasts and Vector Sensors verses Grid Sensors is left to future studies.
1. **Collection** - data is extracted from observed objects
2. **Encoding** - the extracted data is encoded into a grid observation
3. **Communication** - the grid observation is sent to python or used by a trained model
## Collection
A Grid Sensor is the Grid Observation analog of a Unity Camera but with some notable differences. The sensor is made up of a grid of identical Box Colliders which designate the "cells" of the grid. The Grid Sensor also has a list of "detectable objects" in the form of Unity gameobject tags. When an object that is tagged as a detectable object is present within a cell's Box Collider, that cell is "activated" and a method on the Grid Sensor extracts data from said object and associates that data with the position of the activated cell. Thus the Grid Sensor is always orthographic:
A Grid Sensor is the Grid Observation analog of a Unity Camera but with some notable differences. The sensor is made up of a grid of identical Box Colliders which designate the "cells" of the grid. The Grid Sensor also has a list of "detectable objects" in the form of Unity GameObject tags. When an object that is tagged as a detectable object is present within a cell's Box Collider, that cell is "activated" and a method on the Grid Sensor extracts data from said object and associates that data with the position of the activated cell. Thus the Grid Sensor is always orthographic:
Just like the Raycasts mentioned earlier, the Grid Sensor can extract any kind of data from a detected object and just like the Camera, the Grid Sensor maintains the spacial relationship between nearby cells that allows one to take advantage of the CNN literature. Thus the Grid Sensor tries to take the best of both sensors and combines them to something that is more expressive.
Just like the Raycasts mentioned earlier, the Grid Sensor can extract any kind of data from a detected object, and just like the Camera, the Grid Sensor maintains the spacial relationship between nearby cells that allows one to take advantage of the computational properties of CNNs. Thus the Grid Sensor tries to take the best of both sensors and combines them to something that is more expressive.
Lets imagine a scenario where an agent is faced with 2 enemies and there are 2 "equipable" weapons somewhat behind the agent. Lets also keep in mind some important properties of the enemies and weapons that would be useful for the agent to know. For simplicity, lets assume enemies represent their health as a percentage (0-100%). Lets also assume that enemies and weapons are the only 2 kind of objects that the agent would see in the entire game.
Let's imagine a scenario where an agent is faced with two enemies and there are two "equipable" weapons somewhat behind the agent. It would be helpful for the agent to know the location and properties of both the enemies as well as the equippable items. For simplicity, let's assume enemies represent their health as a percentage (0-100%). Also assume that enemies and weapons are the only two kinds of objects that the agent would see in the entire game.
If a raycast hits an object, not only could we get the distance (normalized by the maximum raycast distance) we would be able to extract its type (enemy vs weapon) and if its an enemy then we could get its health (e.g., .6).
If a raycast hits an object, not only could we get the distance (normalized by the maximum raycast distance) we would be able to extract its type (enemy vs weapon) and any attribute associate with it (e.g. an enemy's health).
There are many ways in which one could encode this information but one reasonable encoding is this:
```
For example, if the raycast hit nothing then this would be represented by `[0, 0, 0, 1]`.
If instead the raycast hit an enemy with 60% health that is 50% of the maximum raycast distance, the data would be represented by `[0, 1, .6, .5]`.
The limitations of raycasts which were presented above are easy to visualize in the below image. The agent is unable to see where the weapons are and only sees one of the enemies. Typically in the MLAgents examples, this situation is mitigated by including previous frames of data so that the agent observes changes through time. However, in more complex games, it is not difficult to imagine scenarios where an agent would not be able to observe important information using only Raycasts.
The limitations of raycasts which were presented above are easy to visualize in the below image. The agent is unable to see where the weapons are and only sees one of the enemies. Typically in the ML-Agents examples, this situation is mitigated by including previous frames of data so that the agent observes changes through time. However, in more complex games, it is not difficult to imagine scenarios where an agent might miss important information using only Raycasts.
Following the same data extraction method presented in the section on raycasts, if a Grid Sensor was used instead of Raycasts or a Camera, then not only would the agent be able to extract the health value of the enemies but it would also be able to encode the relative positions of those objects as is done with Camera. Additionally, as the texture of the objects is not used, this data can be collected without rendering the scene.
float[] channelValues = new float[ChannelDepth.Length]; // ChannelDepth.Length = 2 in this example
channelValues[0] = type_index; // this is the observation collected in default implementation
if (currentColliderGo.tag == "enemy")
{
var enemy = currentColliderGo.GetComponent<EnemyClass>();
channelValues[1] = enemy.health; // the value may have to be normalized depends on the type of GridSensor encoding you use (see sections below)
}
return channelValues;
}
```
At the end of the Collection phase, each cell with an object inside of it has `GetObjectData` called and the returned values (named `channelValues`) is then processed in the Encoding phase which is described in the next section.
At the end of the Collection phase, each cell with an object inside of it has `GetObjectData` called and the returned values is then processed in the Encoding phase which is described in the next section.
The CountingGridSensor builds on the GridSesnor to perform the specific job of counting the number of object types that are based on the different detectable object tags. The encoding and is meant to exploit a key feature of the Grid Sensor. In both the Channel and the Channel Hot DepthTypes, the closest detectable object, in relation to the agent, that lays within a cell is used for encoding the value for that cell. In the CountingGridSensor, the number of each type of object is recorded and then normalized according to a max count, stored in the ChannelDepth.
The CountingGridSensor builds on the GridSensor to perform the specific job of counting the number of object types that are based on the different detectable object tags. The encoding is meant to exploit a key feature of the GridSensor. In original GridSensor, only the closest detectable object, in relation to the agent, that lies within a cell is used for encoding the value for that cell. In the CountingGridSensor, the number of each type of object is recorded and then normalized according to a max count.
An example of the CountingGridSensor can be found below.
In order to support different ways of representing the data extracted from an object, multiple "depth types" were implemented. Each has pros and cons and, depending on the use-case of the Grid Sensor, one may be more beneficial than the others.
The data stored that is extracted during the *Collection* phase, and stored in `channelValues`, may come from different sources. For instance, going back the Enemy/Weapon example in the previous section, an enemy's health is continuous whereas the object type (enemy or weapon) is categorical data. This distinction is important as categorical data requires a different encoding mechanism than continuous data.
The stored data that is extracted during the *Collection* phase may come from different sources, and thus be of a different nature. For instance, going back to the Enemy/Weapon example in the previous section, an enemy's health is continuous whereas the object type (enemy or weapon) is categorical data. This distinction is important as categorical data requires a different encoding mechanism than continuous data.
The Grid Sensor handles this distinction with 4 properties that define how this data is to be encoded:
* DepthType - Enum signifying the encoding mode: Channel, ChannelHot
* ObservationPerCell - the total number of values that are in each cell of the grid observation
* ChannelDepth - int[] describing the range of each data within the `channelValues`
* ChannelOffset - int[] describing the number of encoded values that come before each data within `channelValues`
The GridSensor handles this distinction with two user defined properties that define how this data is to be encoded:
The ChannelDepth and the DepthType are user defined and gives control to the developer to how they can encode their data. The ChannelDepth and ChannelOffset are both initialized and used in different ways depending on the ChannelDepth and the DepthType.
* DepthType - Enum signifying the encoding mode: Channel, ChannelHot
* ChannelDepth - `int[]` describing the range of each data and is used differently with different DepthType
How categorical and continuous data is treated is different between the different DepthTypes as will be explored in the sections below. The sections will use an on-going example similar to example mentioned earlier where, within a cell, the sensor observes: `an enemy with 60% health`. Thus the cell contains 2 kinds of data: categorical data (object type) and the continuous data (health). Additionally, the order of the observed tags is important as it allows one to encode the tag of the observed object by its index within list of observed tags. Note that in the example, the observed tags is defined as ["weapon", "enemy"].
How categorical and continuous data is treated is different between the different DepthTypes as will be explored in the sections below. The sections will use an on-going example similar to the example mentioned earlier where, within a cell, the sensor observes: `an enemy with 60% health`. Thus the cell contains two kinds of data: categorical data (object type) and the continuous data (health). Additionally, the order of the observed tags is important as it allows one to encode the tag of the observed object by its index within the list of observed tags. Note that in the example, the observed tags is defined as ["weapon", "enemy"].
The Channel Based Grid Observations is perhaps the simplest in terms of usability and similarity with other machine learning applications. Each grid is of size WxHxC where C is the number of channels. To distinguish between categorical and continuous data, one would use the ChannelDepth array to signify the ranges that the values in the `channelValues` array could take. If one sets ChannelDepth[i] to be 1, it is assumed that the value of `channelValues[i]` is already normalized. Else ChannelDepth[i] represents the total number of possible values that `channelValues[i]` can take.
The Channel Based Grid Observations represent obsevations in a normalized form with 0 to 1. To distinguish between categorical and continuous data, one would use the ChannelDepth array to signify the ranges that the values in the `channelValues` array could take. If one sets ChannelDepth[i] to be 1, it is assumed that the value of `channelValues[i]` is already normalized. Else ChannelDepth[i] represents the total number of possible values that `channelValues[i]` can take and will be used for normalization.
Using the example described earlier, if one was using Channel Based Grid Observations, they would have a ChannelDepth = {2, 1} to describe that there are two possible values for the first channel and the 1 represents that the second channel is already normalized.
As the "enemy" is in the second position of the observed tags, its value can be normalized by:
For continuous data, you should specify `ChannelDepth[i]` to 1 and the collected data should be already normalized by its min/max range. For discrete data, you should specify `ChannelDepth[i]` to be the total number of possible values, and the collected data should be an integer value within range of `ChannelDepth[i]`.
Using the example described earlier, if one was using Channel Based Grid Observations, they would have a ChannelDepth = {2, 1} to describe that there are two possible values for the first channel (ObjectType) and the 1 represents that the second channel (EnemyHealth) is continuous and should be already normalized.
For ObjectType, "weapon", "enemy" will be represented respectively as:
num = detectableObjects.IndexOfTag("enemy")/ChannelDepth[0] = 2/2 = 1;
`[1, .6]`. If the health in the game is not represented in a normalized form, for example if the health is represented in an integer ranging from -100 to 100, you'll need to manully nomalize it during collection. That is, If you get value 50, you need to normalize it by `50/(100- (-100))=0.25` and collect 0.25 instead of 50.
The Channel Hot DepthType generalizes the classic OneHot encoding to differentiate combinations of different data. Rather than normalizing the data like in the Channel Based section, each element of `channelValues` is represented by an encoding based on the ChannelDepth. If ChannelDepth[i] = 1, then this represents that `channelValues[i]` is already normalized (between 0-1) and will be used directly within the encoding. However if ChannelDepth[i] is an integer greater than 1, then the value in `channelValues[i]` will be converted into a OneHot encoding based on the following:
The Channel Hot DepthType generalizes the classic OneHot encoding to differentiate combinations of different data. Rather than normalizing the data like in the Channel Based section, each element of `channelValues` is represented by an encoding based on the ChannelDepth. If ChannelDepth[i] = 1, then this represents that `channelValues[i]` is already normalized (between 0-1) and will be used directly within the encoding which is same as with Channel Based. However if ChannelDepth[i] is an integer greater than 1, then the value in `channelValues[i]` will be converted into a OneHot encoding based on the following:
```
float[] arr = new float[ChannelDepth[i] + 1];
The encoding of each channel is then concatenated together. Clearly using this setup allows the developer to be able to encode values using the classic OneHot encoding. Below are some different variations of the ChannelDepth which create different encodings of the example:
##### ChannelDepth = {3, 1}
The first element, 3, signifies that there are 3 possibilities for the first channel and as the "enemy" is 2nd in the detected objects list, the "enemy" in the example is encoded as `[0, 0, 1]` where the first index represents "no object". The second element, 1, signifies that the health is already normalized and, following the table, is used directly. The resulting encoding is thus:
The first element, 3, signifies that there are three possibilities for the first channel and as the "enemy" is 2nd in the detected objects list, the "enemy" in the example is encoded as `[0, 0, 1]` where the first index represents "no object". The second element, 1, signifies that the health is already normalized and, following the table, is used directly. The resulting encoding is thus:
```
[0, 0, 1, 0.6]
```
### CountingGridSensor
As introduced above, the CountingGridSensor inherits from the GridSensor for the sole purpose of counting the different objects that lay within a cell. In order to normalize the counts so that the grid can be properly encoded as PNG, the ChannelDepth is used to represent the "maximum count" of each type. For the working example, if the ChannelDepth is set as {50, 10}, which represents that the maximum count for objects with the "weapon" and "enemy" tag is 50 and 10, respectively, then the resulting data would be:
As mentioned above, the CountingGridSensor inherits from the GridSensor for the sole purpose of counting the different objects that lay within a cell. In order to normalize the counts so that the grid can be properly encoded as PNG, the ChannelDepth is used to represent the "maximum count" of each type. For the working example, if the ChannelDepth is set as {50, 10}, which represents that the maximum count for objects with the "weapon" and "enemy" tag is 50 and 10, respectively, then the resulting data would be:
At the end of the Encoding phase, all of the data for a Grid Observation is placed into a float[] referred to as the perception buffer. Now the data is ready to be sent to either the python side for training or to be used by a trained model within Unity. This is where the Grid Sensor takes advantage of 2D textures and the PNG encoding schema to reduce the number of bytes that are being sent.
At the end of the Encoding phase, all the Grid Observations will be sent to either the python side for training or to be used by a trained model within Unity. Since the data format is similar to images collected by Camera Sensors, Grid Observations also have the CompressionType option to specify whether to send the data directly or send in PNG compressed form for better communication efficiency.
The 2D texture is a Unity class that encodes the colors of an image. It is used for many ways through out Unity but it has 2 specific methods that the Grid Sensor takes advantage of:
`SetPixels` takes a 2D array of Colors and assigns the color values to the texture.
`EncodeToPNG` returns a byte[] containing the PNG encoding of the colors of the texture.
Together these 2 functions allow one to "push" a WxHx3 normalized array to a PNG byte[]. And indeed, this is how the Camera Sensor in Unity MLAgents sends its data to python. However, the grid sensor can have N channels so there needs to be a more generic way to send the data.
The core idea behind how a Grid Observation is encoded is the following:
1. split the channels of a Grid Observation into groups of 3
2. encode each of these groups as a PNG byte[]
3. concatenate all byte[] and send the combined array to python
4. reconstruct the Grid Observation by splitting up the array and decoding the sections
Once the bytes are sent to python, they are then decoded and used as a tensor of the correct shape within the mlagents python codebase.
Once the bytes are sent to Python, they are then decoded and provided as a tensor of the correct shape.
This version of the Unity ML-Agents Extensions package is compatible with the
following versions of the Unity Editor:
- 2018.4 and later
- If using the `InputActuatorComponent`
- 2019.4 or later
- install the `com.unity.inputsystem` package version `1.1.0-preview.3` or later.
- Else 2018.4 and later
none
- For the `InputActuatorComponent`
- Limited implementation of `InputControls`
- No way to customize the action space of the `InputActuatorComponent`
## Need Help?
The main [README](https://github.com/Unity-Technologies/ml-agents/tree/release_12_docs/README.md) contains links for contacting the team or getting support.
| `curiosity -> strength` | (default = `1.0`) Magnitude of the curiosity reward generated by the intrinsic curiosity module. This should be scaled in order to ensure it is large enough to not be overwhelmed by extrinsic reward signals in the environment. Likewise it should not be too large to overwhelm the extrinsic reward signal. <br><br>Typical range: `0.001` - `0.1` |
| `curiosity -> encoding_size` | (default = `64`) Size of the encoding used by the intrinsic curiosity model. This value should be small enough to encourage the ICM to compress the original observation, but also not too small to prevent it from learning to differentiate between expected and actual observations. <br><br>Typical range: `64` - `256` |
| `curiosity -> network_settings` | Please see the documentation for `network_settings` under [Common Trainer Configurations](#common-trainer-configurations). The network specs used by the intrinsic curiosity model. The value should of `hidden_units` should be small enough to encourage the ICM to compress the original observation, but also not too small to prevent it from learning to differentiate between expected and actual observations. <br><br>Typical range: `64` - `256` |
| `curiosity -> learning_rate` | (default = `3e-4`) Learning rate used to update the intrinsic curiosity module. This should typically be decreased if training is unstable, and the curiosity loss is unstable. <br><br>Typical range: `1e-5` - `1e-3` |
### GAIL Intrinsic Reward
| `gail -> strength` | (default = `1.0`) Factor by which to multiply the raw reward. Note that when using GAIL with an Extrinsic Signal, this value should be set lower if your demonstrations are suboptimal (e.g. from a human), so that a trained agent will focus on receiving extrinsic rewards instead of exactly copying the demonstrations. Keep the strength below about 0.1 in those cases. <br><br>Typical range: `0.01` - `1.0` |
| `gail -> demo_path` | (Required, no default) The path to your .demo file or directory of .demo files. |
| `gail -> encoding_size` | (default = `64`) Size of the hidden layer used by the discriminator. This value should be small enough to encourage the discriminator to compress the original observation, but also not too small to prevent it from learning to differentiate between demonstrated and actual behavior. Dramatically increasing this size will also negatively affect training times. <br><br>Typical range: `64` - `256` |
| `gail -> network_settings` | Please see the documentation for `network_settings` under [Common Trainer Configurations](#common-trainer-configurations). The network specs for the GAIL discriminator. The value of `hidden_units` should be small enough to encourage the discriminator to compress the original observation, but also not too small to prevent it from learning to differentiate between demonstrated and actual behavior. Dramatically increasing this size will also negatively affect training times. <br><br>Typical range: `64` - `256` |
| `gail -> learning_rate` | (Optional, default = `3e-4`) Learning rate used to update the discriminator. This should typically be decreased if training is unstable, and the GAIL loss is unstable. <br><br>Typical range: `1e-5` - `1e-3` |
| `gail -> use_actions` | (default = `false`) Determines whether the discriminator should discriminate based on both observations and actions, or just observations. Set to True if you want the agent to mimic the actions from the demonstrations, and False if you'd rather have the agent visit the same states as in the demonstrations but with possibly different actions. Setting to False is more likely to be stable, especially with imperfect demonstrations, but may learn slower. |
| `gail -> use_vail` | (default = `false`) Enables a variational bottleneck within the GAIL discriminator. This forces the discriminator to learn a more general representation and reduces its tendency to be "too good" at discriminating, making learning more stable. However, it does increase training time. Enable this if you notice your imitation learning is unstable, or unable to learn the task at hand. |
| `rnd -> strength` | (default = `1.0`) Magnitude of the curiosity reward generated by the intrinsic rnd module. This should be scaled in order to ensure it is large enough to not be overwhelmed by extrinsic reward signals in the environment. Likewise it should not be too large to overwhelm the extrinsic reward signal. <br><br>Typical range: `0.001` - `0.01` |
| `rnd -> encoding_size` | (default = `64`) Size of the encoding used by the intrinsic RND model. <br><br>Typical range: `64` - `256` |
| `rnd -> network_settings` | Please see the documentation for `network_settings` under [Common Trainer Configurations](#common-trainer-configurations). The network specs for the RND model. |
| `curiosity -> learning_rate` | (default = `3e-4`) Learning rate used to update the RND module. This should be large enough for the RND module to quickly learn the state representation, but small enough to allow for stable learning. <br><br>Typical range: `1e-5` - `1e-3`
- LSTM does not work well with continuous actions. Please use
discrete actions for better results.
- Since the memories must be sent back and forth between Python and Unity, using
too large `memory_size` will slow down training.
- Adding a recurrent layer increases the complexity of the neural network, it is
recommended to decrease `num_layers` when using recurrent.
- It is required that `memory_size` be divisible by 2.
Ervin Teng
4 年前