# Learning Agents / Getting Started

**URL:** <https://forums.unrealengine.com/t/learning-agents-getting-started/1297712>\
**Category:** Programming & Scripting\
**Tags:** unreal-engine, Learning-Agents, question\
**Created:** [September 15, 2023, 7:59pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712 "2023-09-15T19:59:00Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![MondragonJeux](https://d3kjluh73b9h9o.cloudfront.net/original/4X/2/8/5/28502927f1d97ba1051b3b47ccfe17233bf41ef0.png) [@MondragonJeux](https://forums.unrealengine.com/u/MondragonJeux)\
**Post date:** [September 15, 2023, 7:59pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712/1 "2023-09-15T19:59:00Z")

</div>

Hi,  
Ok, so I’m very excited about Learning Agents and just went through the Quickstart available in the tutorials section, everything seems to be working except that a lot of my agents are training while moving backwards, so it feels something is off. Has anyone tried Learning Agents yet?  
Cheers,

---

<div class="post-metadata">

**Author:** ![Deathcalibur](https://d1ap1mz92jnks1.cloudfront.net/user_avatar/forums.unrealengine.com/deathcalibur/32/1037610_2.png) [@Deathcalibur](https://forums.unrealengine.com/u/Deathcalibur)\
**Post date:** [September 15, 2023, 8:06pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712/2 "2023-09-15T20:06:05Z")

</div>

There are a lot of discussion in this thread. It’s normal during the beginning of training for the cars to go in reverse.

> [@Tutorial: Learning to Drive](https://forums.unrealengine.com/t/tutorial-learning-to-drive/1275490/32):
>
> Hi Mellos, in my experience, this is normal for the first few iterations, as the cars seem to just be doing ‘random inputs’ and whichever results in the highest score gets reinforced; the cars need to ‘learn’ that ‘forward gets you more points’

After a couple hours things should look good. If they don’t improve, then let me know and I will investigate.

---

<div class="post-metadata">

**Author:** ![MondragonJeux](https://d3kjluh73b9h9o.cloudfront.net/original/4X/2/8/5/28502927f1d97ba1051b3b47ccfe17233bf41ef0.png) [@MondragonJeux](https://forums.unrealengine.com/u/MondragonJeux)\
**Post date:** [September 15, 2023, 8:11pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712/3 "2023-09-15T20:11:30Z")

</div>

Yeah, I’m training as I type this, the docs do say it needs a couple hours, I’ll just let it go then. I’m intrigued by the challenge of coming up with better learning rules to achieve “realistic” paths, as apparently the simple implementation they show is not the best. Gonna play with this and digest, I love the idea of being able to build whatever kind of neural network one could think, without leaving the Engine. Have you found something exciting or cool?  
Best,

---

<div class="post-metadata">

**Author:** ![Deathcalibur](https://d1ap1mz92jnks1.cloudfront.net/user_avatar/forums.unrealengine.com/deathcalibur/32/1037610_2.png) [@Deathcalibur](https://forums.unrealengine.com/u/Deathcalibur)\
**Post date:** [September 15, 2023, 8:24pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712/4 "2023-09-15T20:24:14Z")

</div>

Well I wrote the tutorial 🙂

I wanted to keep the tutorial as simple as possible but we’ve done some tests and can definitely get better racing lines to develop.

Right now you’re restricted to only Feed-forward MLPs (if you are familiar with NN architectures). We’re working on expanding that functionality though.

Good luck.

---

<div class="post-metadata">

**Author:** ![MondragonJeux](https://d3kjluh73b9h9o.cloudfront.net/original/4X/2/8/5/28502927f1d97ba1051b3b47ccfe17233bf41ef0.png) [@MondragonJeux](https://forums.unrealengine.com/u/MondragonJeux)\
**Post date:** [September 15, 2023, 8:55pm UTC](https://forums.unrealengine.com/t/learning-agents-getting-started/1297712/5 "2023-09-15T20:55:16Z")

</div>

Oh wow! Well, thank you for an excellent tutorial and even better framework!

So just in general terms how did you go about getting better paths?  
I understand that right now we’re only trying to get the fastest lap possible, but I’d also like to take into account, for instance, the length of the path found that completed an entire lap and try to optimize for that?  
I’ve been going down NNs lately just to get a better understanding of how all of the new AI tools work, but I think it’s enough to say that I have no idea what FFMLPs are 🙂

It would be interesting to see approaches on how to devise strategies or “behaviours” that then are implemented using other functions.

Congratulations and thanks for putting this out for everyone 🤓
