Designing Autonomous AI

Chapter 6. Setting Goals for Your AI Brain

Goals come in pairs.

In order to discover and set goals for your AI, you must understand that there are always going to be trade-offs. I learned this by interviewing hundreds of subject matter experts about their systems and processes, but one particular example cemented it for me. A group of optimization experts asked me about the differences between DRL, gradient descent algorithms, and evolutionary algorithms for industrial decision-making. I explained the differences (see Chapter 2) and then they told me about a problem that they were working on with an industrial process.

The process manages a chemical reaction that produces a valuable chemical product. The price of this product fluctuates significantly with market demand. The catalyst that drives the chemical reaction is also expensive and fluctuates with market demand. They have different goals and strategies, and how they deploy those strategies depends on the market.

There’s Always a Trade-off

Here’s the first goal/strategy pair: When the demand for the chemical is high, its price is also high. So at times when the price of the catalyst is relatively low, the goal is to make as much chemical as possible and the strategy is to consume catalyst as quickly as possible.

Then I asked the process expert about the opposing strategy, the companion strategy that trades off against this one. This surprised one person in the meeting, who told me that the optimal strategy was always to ...