Next generation AI-enabled simulation models are here; maximise your results with Pathmind

Find out how to combine deep reinforcement learning with AnyLogic software to get more from your simulation models

Until now, there hasn’t been an off-the-shelf tool you can use to easily incorporate the benefits of using Artificial Intelligence (AI) with simulation. Pathmind, a brand new SaaS tool for model builders, helps you work out the best way to improve your processes and systems using deep reinforcement learning. The result is a powerful new software combination for any business analyst.

 

Presenting this approach to an AI audience for the first time, DSE Consulting used the meeting venue—a coffee shop—for inspiration. This simple real-world example of a bustling cafe shows you how an AI tool can work out how to improve business operations faster than any human and make fewer mistakes along the way.

How do you make more profit from a coffee shop?

  • Reduce waiting times
     

  • Serve more customers
     

  • Seat more customers
     

  • Keep the seating area clean and tidy

Let’s say that in our imaginary coffee shop customers often have to wait to be served. They wait a second time to pay the bill, so the queues at the counter get longer and longer. Once served, people struggle to find a seat and, as the shop becomes overcrowded, breakages and spillages happen more often which puts off new customers from coming in. As a result, profits don’t rise.

 

How do you fix this?

Without simulation, managers would be forced to make physical changes to the shop floor. Such as trying out a new system to process orders faster, adding more tables to seat more customers, or changing the cleaning routine so that the shop doesn’t become cluttered and dirty. But if those changes don’t then help to increase profits, the time and resources you spent on trialling them are wasted.

 

When you build an animated simulation model of your shop instead, you can see what happens—virtually—when you make a change. Then you compare which changes bring in the biggest increase in profits:

  • You could add a second waiter to take payment and see if that reduces the queues. 
     

  • Introduce a trolley for customers and staff to put dirty dishes on so they’re swiftly moved away from the tables. (Where you place the trolley could be important too, to wheel it quickly into the kitchen without weaving through customers.)
     

  • Moving all of the tables away from the entrance might also help to ease congestion at busy times, so the shop feels less cluttered and there is less chance of somebody knocking over a drink with their elbow.

Testing these kinds of changes brings you clarity, to understand which arrangement works best for your shop. This phase of analysis is traditionally performed by the simulation professional, but if you build simulation models using AnyLogic software, you can now add an AI component to do this for you. You can then:

  • Get your results faster since the AI works quicker than you do

  • Work on much bigger projects because a machine can handle more complicated data sets

  • Trust that your data is more accurate by removing human error

What is simulation modelling?

 

Simulation modelling allows you to test the unknown. 

 

You can build a digital copy of a real-world environment, or create a new one — even if it doesn’t exist yet. Then you can run tests in your virtual world, not on your live operation, so there’s no interruption to service.

 

  • Use the AnyLogic software to analyse your results 

  • Make predictions about your future

  • Validate your ideas (and get funding to pursue them)

DSE Consulting uses AnyLogic software because there are three types of simulation modelling and AnyLogic is the only product to combine them all. No having to choose one simulation technique over another or to use different software for different tasks.

What is reinforcement learning?

 

 

“We don’t learn to walk by following rules. We learn by doing and by falling over.” — Richard Branson

Reinforcement learning is a type of machine learning which falls under the broader category of AI. It simply reinforces good behaviour using penalties and rewards.

Consider training a new waiter to clear tables in a real coffee shop. They get paid for doing this successfully (a reward) but would lose their job if they continuously broke cups and saucers (a penalty). If they want to stay employed, they have to learn how to stack the dirty cups and then how to transport them safely back to the kitchen for cleaning so there’s less risk of being fired. 

 

When a machine can learn how to get better at something, it's said to have artificial intelligence.

With the introduction of reinforcement learning the machine begins to learn what makes the stick fall, and how to avoid it, which keeps the stick upright for longer.

Fully trained, the machine learns to balance the stick perfectly upright, all of the time ― much faster than you or I can learn to balance a pencil on the palm of our hand.

Want to take on the machine?

 

Pick up a pencil and try to balance it on the palm of your hand. Keep the pencil upright without it falling over.

 

If you’re having trouble, you’re not alone!

 

Here’s a machine trying to balance a stick with no training. The stick falls immediately:

Deep reinforcement learning (DRL)


 

Deep learning adds a new dimension to reinforcement learning because the machine does more than learn through trial and error. It applies what it learns to create new routes to success. In human terms, you could consider it a form of thinking for itself.

Now with Pathmind, for the first time, you can apply DRL to your simulation model. This is the closest we’ve come to having a machine optimise a simulation model in much the same way as people do to find the quickest, best, or most profitable way to operate your business.

How does DRL help our coffee shop?

In the simulation coffee shop we’ve mocked up for this demonstration, the waiter can do four things: 

 

  1. Clean tables

  2. Take orders

  3. Make the coffee

  4. Take payment

 

He gets rewards for success. For example, the amount of money he takes, a shortened waiting time for customers, or the number of orders processed. 

 

In the first video below, the waiter completes his four tasks in random order. As you can see from the charts that appear, the result isn’t great. The pie chart shows in green that only 45% of the customers are seated and served successfully. 

Traditionally, you would now continue to make changes to the model and run further tests until you see an improvement in your results. But now, you can import your model to Pathmind, allow the AI to do its work, and export it back to AnyLogic to inspect your results once it’s done. 

 

You save time, remove the possibility of human error, and Pathmind can handle much bigger data sets than our human brains can cope with. 

Trained simulation model using Pathmind

 

In the second video, the Pathmind AI tool now starts to train the waiter in the simulation model. The waiter starts to learn that keeping the tables clean and processing orders quickly gives him the most rewards. Already, the wait times are shorter and more people are served, so the number of satisfied customers rises to 65-80%. That’s a significant improvement to your operation in less time than it would take for you to achieve these results without using AI.

Through deep reinforcement learning, the waiter will now continue to better his performance on each of the four tasks. The goal is to work out how to get the shortest waiting time, the highest number of orders taken, and the most money in the till. When you export your simulation model from Pathmind back to AnyLogic, the results will show you the best way to maximise your return on running your coffee shop so you can take immediate action on implementing worthwhile changes.

Want to see what Pathmind can do for you? Book a demo

When DRL goes wrong...

 

Of course, you have to train the AI correctly. This is what happens if there’s no penalty given for a large queue forming. The waiter doesn’t try to find a way to reduce the crowd because he has no idea it’s a bad thing.

What's next for simulation and AI?

 

Developing deep reinforcement learning tools for use with simulation models is a natural next step for the industry. Business analysts get faster, more accurate results, which means you make better decisions about the next steps for your business — even for multinational, complex systems with many layers to the organisation.

If you'd like to see how simulation modelling is used in your industry or to find out more about using AI, we'll show you.

Book a demo today

© 2019 DSE Consulting, all rights reserved