In my [last post](/machine-learning/a-game-developer-learns-machine-learning-a-little-deeper/) we dug a little deeper into machine learning by starting from scratch and training an agent on a vastly simplified version of mr nibbles. In this post we will explore what happens when we step it up a notch and attempt to train on a significantly more complex version of the game.



# The Series

If you are new to the series I recommend you checkout the previous posts first, here's the whole series:

0. [A Game Developer Learns Machine Learning - Intent](/machine-learning/a-game-developer-learns-machine-learning-intent/)
1. [A Game Developer Learns Machine Learning - Getting Started](/machine-learning/a-game-developer-learns-machine-learning-getting-started/)
1. [A Game Developer Learns Machine Learning - A Little Deeper](/machine-learning/a-game-developer-learns-machine-learning-a-little-deeper/)
1. A Game Developer Learns Machine Learning - Mr Nibbles Basics

# Code

If you want to see the code for this post then checkout this tag: https://github.com/mikecann/MrNibblesML/tree/MrNibblesBasicsBlogPost

# Mr Nibbles Forever vs Mr Nibbles

My original [Intent](/machine-learning/a-game-developer-learns-machine-learning-intent/) was to teach an ML agent to learn my game [Mr Nibbles Forever](http://epicshrimp.com/app/mrnibblesforever/). Based on my experience up to this point I think what might be easier is to begin with trying to train an agent to learn a game similar to "[Mr Nibbles](http://mr-nibbles.com/)" instead of "[Mr Nibbles Forever](http://epicshrimp.com/app/mrnibblesforever/)".

<iframe width="853" height="480" src="https://www.youtube.com/embed/lyAf7VVLdKg" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

Rather than an endless procedurally generated game like Mr Nibbles Forever, Mr Nibbles is a level-based platform-puzzler.

I suspect that Mr Nibbles will be easyier to learn as it's significantly less complex than its later variant, Mr Nibbles Forever.

# The Environment

Before I can even think about training my agent, I need an environment. Mr Nibbles was originally made many years ago in HaXe and thus I will need a way to build levels in Unity.

Unity recently launched a [set of tools](https://docs.unity3d.com/Manual/Tilemap-Painting.html) for generating tile-based games.

So I set to work using those tools to construct some levels:

<iframe width="853" height="480" src="https://www.youtube.com/embed/99TRpAmZFYM" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

I borrowed some code from Unity's [tutorial on platform games](https://unity3d.com/learn/tutorials/topics/2d-game-creation/creating-basic-platformer-game) to build a "Platform Controller" for Mr Nibbles. This is the thing that actually controls the how mr nibbles moves given an input.

I added a little of flourish in the way of big tubes to represent the entrance and exit points for the level. I even wrote some code that animated Mr Nibbles through the tubes but I found that it caused issues in the training so I decided to drop that aspect for now, perhaps ill add it again later.

The end result is that you are able to play a number of levels as mr nibbles. When you reach the exit another level is randomly selected and Mr Nibbles is spawned into that.

<iframe width="853" height="480" src="https://www.youtube.com/embed/IHdPzaLRY98" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

# The Agent

With the levels built its time to code up the agent.

## State

What information do we need to tell the network about the state of the world that will allow it to successfully learn how to navigate it?

Well we obviously need to supply the position of Mr Nibbles and the position of the exit point.

```csharp
public override List<float> CollectState()
{
    var state = new List<float>();

    state.Add(_exitPoint.transform.position.x);
    state.Add(_exitPoint.transform.position.y);

    state.Add(_platformController.transform.position.x);
    state.Add(_platformController.transform.position.y);

    return state;
}
```

We should probably also supply a variable to help with learning the potential set of actions depending if the player is in the air or not.

```csharp
public override List<float> CollectState()
{
    var state = new List<float>();

    ...

    state.Add(_player.IsGrounded ? 1 : 0);

    return state;
}
```

Im not sure if this is really needed but I thought it cant hurt.

Now for the hard part. How do I represent the level? I cant really pass in the collision bounds for each tile as that would be too complicated to represent as floats.

Each level is constructed as tiles, so everything sits on a uniform grid. So I should only need to pass in the centre point of each tile in the level along with its type and hopefully that should be enough for the agent from.

[![](./tile-xy-pos.jpg)](./tile-xy-pos.jpg)

The next issue is; how do I deal with different level sizes? [This GitHub issue](https://github.com/Unity-Technologies/ml-agents/issues/139) mentions that the state size needs to be fixed (I believe this is to do with the way TensorFlow works under the hood).

This restriction is going to make things a little tricky because our levels are different sizes with a different number of tiles, so keeping the state size the same for each level may be a problem (more on potential solutions to that later).

For now my "just get it working" solution is just to make sure the levels are quite small then grab all the tiles in a certain bounds. If the tile is empty I set its type to "0" if its a solid tile its "1" and if its a spider its "2".

[![](./tile-types-rendered.jpg)](./tile-types-rendered.jpg)

When the level changes I grab the tiles from the tilemaps:

```csharp
public BoundsInt tileBoundsToIncludeInState = new BoundsInt(-25, -1, 0, 40, 1, 1);

public override void AgentReset()
{
    ...
    _tiles = _game.CurrentLevel.GetComponentInChildren<TilesController>().GetTiles(tileBoundsToIncludeInState);
    ...
}
```

Then feed them into the state list:

```csharp
public override List<float> CollectState()
{
    ...

    foreach (var tile in _tiles)
    {
        state.Add(tile.position.x);
        state.Add(tile.position.y);
        state.Add(tile.type);
    }

    ...
}
```

## Actions

Now we have our state, we need to allow the agent to take and action and receive a reward or punishment for that action.

```csharp
public override void AgentStep(float[] actions)
{
    PerformActions(actions);
    UpdateRewards();
}
```

Digging into `PerformActions` first:

```csharp
public const int None = 0;
public const int MoveLeft = 1;
public const int MoveRight = 2;
public const int Jump = 3;

...

private void PerformActions(float[] actions)
{
    var nothing = (int) actions[None] == 1;
    var moveLeft = (int) actions[MoveLeft] == 1;
    var moveRight = (int) actions[MoveRight] == 1;
    var jump = (int) actions[Jump] == 1;

    var isJumping = false;
    var hozMove = 0f;

    if (moveLeft)
        hozMove = -1f;
    if (moveRight)
        hozMove = 1f;
    if (jump)
        isJumping = true;

    _platformController.Tick(hozMove, isJumping);
}
```

I handle the condition where the player can perform none, one or all of the actions simultaneously. I then pass those actions to the platform controller component which updates the physics of mr nibbles.

Now we need to handle how to reward the agent for its actions:

```csharp
private void UpdateRewards()
{
    if (_exitPoint.IsTriggered)
    {
        Wins++;
        reward = 10;
        done = true;
    }
    else if (_spiders.IsTriggered)
    {
        Deaths++;
        reward = -10;
        done = true;
    }
    else
    {
        reward = -0.01f;
    }
}
```

The agent is rewarded for reaching the exit, punished for touching the spiders and given a small punishment for each moment that goes by that it doesnt reach the exit.

Im not sure if these are the correct values, I played around with a few different values but im not sure if these are correct or not.

# Defining Brain and Academy

[As with my last post](/machine-learning/a-game-developer-learns-machine-learning-a-little-deeper/) I am going to leave the academy empty for now, I dont need its features as I dont have multiple agents in my scene.

The brain is setup like so:

[![](./brain-setup.png)](./brain-setup.png)

We have defined the number of states: tiles (x,y,type) + agent state (x,y,isOnGround) + exit point (x,y) = 1505

The number of actions: jump, left, right, nothing (stand still).

The action space is set to continuous because we want to allow the agent to perform multiple actions at once, jump and move right for example.

The state space is continuous because we arent dealing with integer positions for things in the level rather we are using floats.

# Training

With the environment constructed lets move onto training. First I need to get the hyperparameters right for the PPO algorithm. I re-read [the unity best practices doc](https://github.com/Unity-Technologies/ml-agents/blob/master/docs/best-practices-ppo.md) and took a guess at the following hyperparams:

```python
### General parameters
max_steps = 5e5 # Set maximum number of steps to run environment.
run_path = "mrnibbles/run14-differentrewards" # The sub-directory name for model and summary statistics
load_model = False # Whether to load a saved model.
train_model = True # Whether to train the model.
summary_freq = 10000 # Frequency at which to save training statistics.
save_freq = 50000 # Frequency at which to save model.
env_name = "mrnibbles" # Name of the training environment file.
curriculum_file = None

### Algorithm-specific parameters for tuning
gamma = 0.99 # Reward discount rate.
lambd = 0.95 # Lambda parameter for GAE.
time_horizon = 1500 # How many steps to collect per agent before adding to buffer.
beta = 1e-3 # Strength of entropy regularization
num_epoch = 6 # Number of gradient descent steps per batch of experiences.
num_layers = 2 # Number of hidden layers between state/observation encoding and value/policy layers.
epsilon = 0.2 # Acceptable threshold around ratio of old and new policy probabilities.
buffer_size = 10000 # How large the experience buffer should be before gradient descent.
learning_rate = 3e-4 # Model learning rate.
hidden_units = 64 # Number of units in hidden layer.
batch_size = 2500 # How many experiences per gradient descent update step.
normalize = False
```

I kicked off training and fired up TensorBoard. After a few different attempts I felt like I had something that was training so I let it run for an hour or so:

[![](./tensorboard.png)](./tensorboard.png)

It looks okay, we can see that the cumulative reward increases sharply and levels off and the episode length decreases as the agent learns to play the level.

When we try our model out in the game however we see this:

<iframe width="853" height="480" src="https://www.youtube.com/embed/S21Rkob0XfM" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

On the positive side, it looks like he is attempting to move towards the exit point when the exit is to the right of the spawn point. On the negative side he doesnt ever seem to want to try going left and also seems to get stuck for a long time at a small hurdle.

I think whats going on is the levels where the exit point is to the left of the spawn point have spiders, thus they are significantly more difficult for the agent to solve than the levels where the exit point is to the right of the spawn point. As a result the agent has learnt that probably it should always just go right and is unable to learn that it can jump over the spiders.

It looks like the agent is having trouble with jumping in general..

# Future Work

I tried a few smaller tweaks but none seemed to have a significant effect. I think what I need to to change my training strategy.

## Curriculum Training

While browsing through the Unity docs I came across the concept of [curriculum training](https://github.com/Unity-Technologies/ml-agents/blob/master/docs/curriculum.md). As I understand it; you incrementally train your agent with progressively more difficult scenarios thus accelerating the learning by focusing the agent on learning just one aspect of your game at a time. This reminds me of how humans learn in school, we aren't immediately thrown into quantum mechanics, instead we start with basic mathematics and work our way up from there.

I think my next set of experiments will have to take advantage of curriculum learning by slowly introducing level elements until the agent is able to successfully proceed.

## Tile Radius State

One current limitation of my current setup is that I encode the state of the entire level. This means I am currently limited to small levels of fixed size.

A solution (suggested by a member of [my local Machine Learning meetup group](https://perthmlgroup.wordpress.com/)) is to only provide the state of tiles that lie in a radius around Mr Nibbles. As Mr Nibbles moves so to does the nearest surrounding tiles. This is analogous to using a camera as my input state but unlike camera state the agent should be able to learn [temporal information](https://github.com/Unity-Technologies/ml-agents/issues/183).

## Reward Tweaking

I think I may need to work on how I reward and punish a little bit more. Perhaps I should have a timeout if it takes the agent too long to solve a level, maybe that would prevent it from learning that on certain levels its best just to suicide rather than try to solve the level.

## Velocity

Another potential issue is was unable to learn to jump correctly was because it was missing information about its velocity. I was hoping it could infer the velocity based on the previous positional state of mr nibbles but perhaps it would help if I manually supplied that information as two more state parameters.

# Conclusion

Well I cant really call this post a success, but I certainty learnt a lot and that's the ultimate goal. I have lots of ideas for how to proceed from here so subscribe to stay tuned.

Until next time!

P.S. If you want to see the code for this post then checkout this tag: https://github.com/mikecann/MrNibblesML/tree/MrNibblesBasicsBlogPost

I mentioned [in my last post](/machine-learning/a-game-developer-learns-machine-learning-intent/) that the intent is to write an AI to play [Mr Nibbles Forever](http://epicshrimp.com/app/mrnibblesforever/) better an any human. Before we can get anywhere near that goal however we need to get the very basics working first.



# The Series

If you are new to the series I recommend you checkout the previous posts first, here's the whole series:

0. [A Game Developer Learns Machine Learning - Intent](/machine-learning/a-game-developer-learns-machine-learning-intent/)
1. A Game Developer Learns Machine Learning - Getting Started
1. [A Game Developer Learns Machine Learning - A Little Deeper](/machine-learning/a-game-developer-learns-machine-learning-a-little-deeper)
1. [A Game Developer Learns Machine Learning - Mr Nibbles Basics](/machine-learning/a-game-developer-learns-machine-learning-mr-nibbles-basics)

# TLDR;

We will use Reinforcement Learning to train an agent to balance a ball.

<iframe width="853" height="480" src="https://www.youtube.com/embed/ZnBfvARKXeo" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

# Unity ML Agents

I find that when tacking something new for the first time, the best thing to do is take something built by others, then hack around with it for a little bit to understand the very basics. So rather than starting from total scratch im going to look for something that already exists.

As luck would have it, Unity have [just released](https://blogs.unity3d.com/2017/09/19/introducing-unity-machine-learning-agents/) a new set of tools to help game developers use machine learing in their games, perfect!

# Getting Setup

First things first, I clone the Unity ML Agents repo https://github.com/Unity-Technologies/ml-agents.

Then I start following their [Getting Started With Balance Ball](https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Getting-Started-with-Balance-Ball.md) docs.

The first thing I need to do is get my system setup with the neccessary tools for machine learning. I follow this guide to do so: https://unity3d.college/2017/10/25/machine-learning-in-unity3d-setting-up-the-environment-tensorflow-for-agentml-on-windows-10/

This process takes quite a while, make sure you follow the instructions EXACTLY. At the bottom of that page is a video which is helpful.

<iframe width="853" height="480" src="https://www.youtube.com/embed/qxicgknzUG8" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

# Jupyter Notebook

So with my system now setup I return back to the [Balance Ball](https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Getting-Started-with-Balance-Ball.md) doc.

The next step is to open Jupyter Notebook which is a popular tool with ML community for training models.

[![](./jupyter-basics.png)](./jupyter-basics.png)

Its actually a pretty aweomse tool. Its sort of like a [REPL](https://en.wikipedia.org/wiki/Read%E2%80%93eval%E2%80%93print_loop) mixed with word document.

You can write code right inside of the doc without firing up and IDE and then execute it and see the results inline.

# Testing the Bridge

Next up is the "Testing Python API" section of the docs.

In Unity the balance ball scene is compiled into an exe, this is whats known as the "environment".

[![](./showing-exe.png)](./showing-exe.png)

Python in the notebook talks to the "brain" in the environment. Its a sort of native bridge between the machine learning API in Python and your "game" / "environment".

[![](./bridge.png)](./bridge.png)

When the training session is started from python, the environment exe is executed which opens as a separate window on your computer. This surprised me as I was expecting it to be hidden background process of some sort.

The environment window that opens up is tiny, presumably so your GFX card doesn't have to waste cycles on rendering and can instead focus on ML.

[![](./env-window.png)](./env-window.png)

The "brain" in the environment is able to receive commands which are called "actions". When an action runs, it changes the "state" of the environment which is returned back to Python. That process of `State > Action > New State` is repeated over and over again (along with some other magic) and out pops our agent.

[![](./learn-loop.png)](./learn-loop.png)

# Training the Model

Now we know our bridge from Pythong to Unity works okay we can move onto actually training the model.

Folling the "Training with PPO" section of the docs we open "TensorBoard" which gives us stats on the the training process as it runs.

[![](./tensorboard1.png)](./tensorboard1.png)

Im not really sure what all those stats mean but thats okay at this stage, we are just tinkering!

The main thing is my model is busy training itself as evidenced by the changing graphs and my pegged CPU 😜

[![](./pegged-cpu.png)](./pegged-cpu.png)

# Testing the Model

Once the "cumulative_reward" is above 75 the model is ready to rock. So I stop the training process in Jupyter.

Theres one last step before we can run the model in Unity, we must add support for TensorFlow directly in Unity in the form of the TensorFlowSharp library.

[![](./tf-sharp.png)](./tf-sharp.png)

TensorFlowSharp requires the experimental .Net 4.6 support in Unity. Thats something to note because as my goal is to eventually get this working on Mr Nibbles Forever which is built on the old .Net 3.5

With the library installed we copy over the generated model into the game and huzzah! It works!

[![](./it-works.png)](./it-works.png)

# Adding some Dynamics

Its awesome that it works but its hard to tell if this is actually balancing the ball dynamically or not. I think if we add a dynamism it will make things clear.

Lets add this script to each of the platforms:

```csharp
using System.Collections;
using System.Collections.Generic;
using UnityEngine;

public class BallNudger : MonoBehaviour
{
    public Color nudgeColor = Color.red;
    public float forceMultiplier = 5f;
    public float colorChangeTime = 0.5f;

    private bool _isNudging = false;
    private Renderer _renderer;
    private Rigidbody _rigidBody;
    private Color _startingColor;
    private float _colorChangeAge;

    void Awake()
    {
        _renderer = GetComponent<Renderer>();
        _rigidBody = GetComponent<Rigidbody>();
        _startingColor = _renderer.material.color;
        _colorChangeAge = colorChangeTime;
    }

    void OnMouseDown()
    {
        _colorChangeAge = 0;
        var ray = Camera.main.ScreenPointToRay(Input.mousePosition);
        RaycastHit hit;
        if (Physics.Raycast(ray, out hit))
            _rigidBody.AddForceAtPosition(ray.direction * forceMultiplier,
                hit.point, ForceMode.Impulse);
    }

    void Update()
    {
        _colorChangeAge += Time.deltaTime;
        _renderer.material.color = _colorChangeAge < colorChangeTime ?
            nudgeColor : _startingColor;
    }

    private Vector3 GetRandomForce()
    {
        return new Vector3(
            Random.Range(-forceMultiplier, forceMultiplier),
            Random.Range(-forceMultiplier, forceMultiplier),
            Random.Range(-forceMultiplier, forceMultiplier));
    }
}
```

Now when we click the ball, it changes color and gets a little nudge. Lets run it again and yes, indeed the AI is able to balance the ball dynamically as the input changes.

<iframe width="853" height="480" src="https://www.youtube.com/embed/ZnBfvARKXeo" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>

# Conclusion

Cool! Well that wasn't as hard as I thought it was going to be.

Sure I haven't actually done any of the hard work or really learnt much about the fundamentals yet but at least I have something working, ready for me to hack with.

So stay tuned, subscribe below as next episode im going to dive a little deeper.

Lately I have spent a considerable amount of time thinking about where I want to go in my career and what sort of things do I want to work on going forward.



# The Series

This is a series of multiple parts, if you liked this post then checkout the others:

0. A Game Developer Learns Machine Learning - Intent
1. [A Game Developer Learns Machine Learning - Getting Started](/machine-learning/a-game-developer-learns-machine-learning-getting-started/)
1. [A Game Developer Learns Machine Learning - A Little Deeper](/machine-learning/a-game-developer-learns-machine-learning-a-little-deeper)
1. [A Game Developer Learns Machine Learning - Mr Nibbles Basics](/machine-learning/a-game-developer-learns-machine-learning-mr-nibbles-basics)

# The History

[![](/wp-content/uploads/2007/09/Guess.jpg)](/wp-content/uploads/2007/09/Guess.jpg)

Before you think about where you want to go next sometimes it helps to think about where you have come from.

When I was 7 years old I [coded my first game](http://localhost:4500/about/). I loved the thrill of turning a meer figment of my imagination into something I could see on the screen and delight when shown to others.

At 18 I went to University and studied Computer Games Programming and loved every second of it, far surpassing my performance during high school, eventually graduating with a first degree with honours.

For the next 10+ years I proceeded to work in the games industry first starting in the flash games world then social games before moving to mobile.

Games are not the only thing I have been interested in over the years as this blog [certainty can attest to](/all-tags/). However games are definitely my core competency.

# The New Shiny

So when you are in an industry that moves so fast with so many new technologies being released every day, it can be really hard to know where to spend your time, particularly for someone like me who "just cant stop tinkering with things" ™.

Machine Learning (ML) is one of the biggest technologies to emerge over the past couple of years. I honestly think that ML has the power to transform every single industry from finance, to agriculture to sports science.

I also think that the games is no different from the others and wont be spared from its transformative effects. Some think that the games industry is safe because its a "creative" industry and thus not susceptible to being replaced by a logic driven machine.

Instead I see a future where an AI is capable of generating an entire game based on your preferences such as "I would like a game that plays like Mario but has the same size and scope as World of Warcraft", the AI then generates variants on that for you to play.

Machine learning opens so many potential possibilities and opportunities for someone with the idea and ability to explore them.

So with that in mind I think that it behove me to at least consider exploring ML a little more.

# The Concern

<div style="width:100%;height:0;padding-bottom:56%;position:relative;"><iframe src="https://giphy.com/embed/rdvtXVH7jDjKU" width="100%" height="100%" style="position:absolute" frameBorder="0" class="giphy-embed" allowFullScreen></iframe></div><p>

But will I enjoy it? This is a major concern of mine. From what I have seen thus far, machine learning is very math heavy, invloves a lot of Python (dynamic programming, _shudders_) and isnt particularly visual. All negatives for someone who isnt great at maths, doesnt really like dynamic languages and is a games programmer.

I am also concerned that this I will basically be starting from scratch again. All my games programming knowledge counts for nothing. Will I be able to adapt to this new set of problems, solutions and tools?

So I think only way to know for sure tho is just to try it and see and the best way to try is to tinker! ™

# Learn to Learn

So over the next few months im going to dig deeper into machine learning. Try to understand it how it works and how it applies to game development.

Im going to blog each step of my adventures. This serves two purposes, firstly I find that writing things down really helps me understand things at a deeper level and secondly I hope that if I dont do a terrible job at explaining it will help others to learn too.

Im going to call this series of blog posts <strong>"A Game Developer Learns Machine Learning"</strong>.

# The Series Goal

[![](https://www.mikecann.blog/wp-content/uploads/2016/05/MrNibblesForever_Gif_5.gif)](https://www.mikecann.blog/wp-content/uploads/2016/05/MrNibblesForever_Gif_5.gif)

The goal for the series is I want to be able to write an AI that is able to play my game [Mr Nibbles Forever](http://epicshrimp.com/app/mrnibblesforever/) better than any other human player.

I have no idea if thats an achievable goal but I think its a goal that should give me a solid understanding of the fundamentals of reinforcement learning and give me a target with which to focus on and guide me through this dense but wonderful topic.

So if you are interested in following along, either [add this blog to your favourite RSS reader](/atom.xml) or enter your email below to subscribe and get email updates.

agdlml

agdlml

agdlml