Taqging based evaluation functions

 Despite its philosophical implication, Artificial Intelligence is also a technical subject within computer science. There are even some algorithms available to realize game playing agents in software. The most famous one is perpahs an evaluation fucntion which is used since the 1980s in computer chess. The inner working of evaluation function including a semantic improvement should be explained next.

In a classical heuristic evaluation function the current game state is mapped towards a score from 0.0 to 1.0 and this score helps the gradient descent solver to find the optimal action. In case of a path planning problem on a 2d grid map, the score is the manhattan distance to the goal, and in case of computer chess the is the strength of a player determined by the amount of chess pieces on the board.

For complex problem from robotics, the score can't be determined directly because its unclear if a certain game state is an improvement or not. Especially for biped walking based on mulitiple joints its impossible to determine a numerical score for a certain pose. What can be utilized instead is a tagging mechanism. Instead of calculating the numerical score, the computer determined only the detected tags. A tag is a semantic anchor like [finger_open] or [battery_low]. Such a tag doesn't describe the game state with numbers but with words. The numerical scores have to be calculated in a second step:

Game state -> semantic tags -> numerical score

The main advantage of a tagging mechanism is that its much easier to encode domain specific knowledge. There is no need to describe a game state from a mathamtical perspective, but its enough to provide a list of words to annotate a game state. This works great easpecially for complex motion capture annotation. An example list of detected tags for a bipod robot might be: [left_foot_front] [balance_stable] [right_foot_back] [servo_motor_off].

The objective for the computer program is to determine the correct tags in realtime. The robot is moving on the screen and in the background all the tags are shown on the screen.