Old school artificial intelligence

Before the advent of large language models, there was lots of AI related research available which was mostly ignored by the public. Typical topics in robotics from the past was NEAT neuroevolution, model predictive control and genetic algorithms. All these concepts were working with computational paradigm. At first the fitness function was defined, e.g. the robot gets +1 reward for moving on step forward, and then an algorithm modifies the parameters of the neural network. In case of model predictive control it was even possible to plan one step ahead, so that the robot was able to use a trial&error strategy.

These concepts can be called old school artificial intelligence because they were working with computers and algorithms as the core element. The assumption was that artificial intelligence is equal to search, so its a mathematical optimization problem. Instead of finding the prime numbers from 1 to 1 million and find 1000 digits of the PI constant, the goal was to use raw processing power to solve robotics problems.

The mentioned NEAT and genetic algorithms strategy were the core principle of AI in the past. Large scale projects like CAM brain and the fifth generation computers in Japan were built around these subjects. The researchers have treated AI problems as mathematical optimization problems by transforming robot motion planning into a mathematical state space first, and then use algorithms to solve these problems.

What was unclear until around 2010 was, that all these strategies have failed. Mathematics and algorithms in general are not powerful to solve robotics problems. This insight is even today some sort of blasphemy because mathematics and algorithms are the building blocks in computer science. If both concepts are rejected as weak then the amount of remaining strategies is small and even empty.

The major result of old school artificial intelligence was not to build robots but to proof that mathematical optimization is a dead end. It can be formulated from a theoretical perspective, that all interesting robotics problems are np hard, and that genetic algorithms including NEAT are not able to solve this problem category. Therefor AI can't be realized.

Its possible to replicate the failure of NEAT and similar approaches with today's software libraries. At first a robotics problem gets formulated, e.g. a walking robot in the OpenAI gym which is a 2d simulator. then a state of the art neuroevolution library is utilized to control the robot. A handcrafted fitness function will reward the algorithm so that the optimization can evolve into improved biped walking generator. Unfortunately, the described setup won't work in reality. Even if the robot walks some steps forward the robot will struggle on larger obstacles. At the same time, the CPU consumption is very high.

At first the fitness function of the robot will improve, but then there is no further improvement available and its impossible to fix the problem. Some reinforcement learning advocates are explaining that with parameter tuning the problem can be solved, other NEAT users claim that with future algorithm update the issue might be solved. But the sad reality is, that mathematical optimization in general fails for robotics problem. No matter if NEAT, genetic algorithm or Q learning was utilized, none of these strategies can control a biped walking robot.

Such kind of insight isn't the end of robotics research but it provides meaningful facts about the limitation of old school algorithms. It shows on a practical example, what sort of AI algorithm isn't working in reality. This opens the path for more advanced AI techniques beyond classical computer science.