Imagine you're a selfish taxi driver who takes passengers around between 3 towns. Each trip has a different probability of catastrophe and different earnings potential. You want to find a repeatable path of trips to maximise profit. No, you can’t just greedily pick the best trip from each city - sometimes, you must forego immediate profit to take a more profitable trip later.

This problem can be framed as a Markov Decision Process. This framing is the building block of reinforcement learning, a field which will likely play a key role in shaping AGI systems. And more importantly, an MDP is one of the simplest models of an agent which we can build upon to learn more about agency in general.

Example: Taxi driver

Let’s describe the taxi driver’s situation using a graph called a Markov chain. For now we’ll ignore the risk of bad things happening on any trips, and just compare the payoffs for taking different routes:

Each town in the graph is known as a “state”. From each state, you can take a couple of actions that involve traveling to a different city. Each action has a different payoff to reflect the current market equilibrium price of different trips. For now, we’ll just pretend that the prices don’t shift over the course of the day, so a taxi driver could always get these payoffs from driving. Now to add some uncertainty into the mix:

That graph looks a bit more complicated. Everywhere we previously had a single arrow, there is now a split arrow. That’s because every trip has two possible outcomes. We could either take a normal trip and make money, or we could have a car crash and end up in the hospital, losing money in the process. Each possible outcome of a decision has a payoff and a transition probability - I just omitted these from the graph for clarity’s sake. For concreteness, let’s zoom into the trip from Berkeley to San Francisco.

Deciding to travel from Berkeley to San Francisco results in:

A 95% chance of an uneventful taxi ride, earning 70 reward
A 5% chance of a car accident and a trip to the hospital, earning -1000 reward

We’ve just created a Markov chain model of the real-world taxi driver situation. These graphs are a bit different from the Markov chains I worked with in a university class. Our Markov chains didn’t have any notion of choices between different actions. An MDP can be viewed as a Markov chain describing the world, with the added fun...

arvganesh

arvganesh

arvganesh

arvganesh