Is contempt factor used as a way to make engine games more interesting?

Question

Chess is believed to be a theoretical draw. So top chess engines, which are not too far off from theoretical optimum, tend to reach drawn positions, which are not too exciting.

My understanding is that to combat this, engine tournaments like TCEC force an "interesting" opening on the engines.

What I don't understand is why this should be necessary, when there is a seemingly more natural way to reach the desired effect - which is to use a contempt factor, i.e. a lower score for draws, for both players.

With high enough contempt, both engines will strive for complex, tactical, interesting positions for which neither is sure who will win, but are unlikely to result in a draw.

With the advent of ML-based engines like Alpha Zero and LC Zero, this is even more natural - the contempt can be baked into the training, so the engine doesn't just avoid draws in-game, the learned network treats drawish positions as inferior.

But I've not seen references to contempt being used that way - rather it appears to be a tool to maximize the traditional scoring, when there is reason to believe a difference in level between the engine and its opponent. Its goal is not to make games more interesting for us human observers.

So - am I correct that contempt is not currently an accepted way to make engine games more interesting? And if so, why not?

Answer 1

Engine competitions don't use contempt to make games more interesting because engines that have high contempt are weaker than ones with low contempt. Put Stockfish (with handcrafted eval) with c = 100 against Stockfish with c = 0, and the latter Stockfish is likely to win. Therefore, if you force all competitors to use high contempt, then you don't actually have the strongest engines playing, and you cannot call your competition the "Top" chess engine championship.

Answer 2

The first use of contempt factor was. I'm fairly certain, to make the engine 'aware' that human opponents are fallible: even if the engine scores the situation as theoretically drawn, it is likely that the opponent can't see or can't maintain that, and so make the engine go for 'drawn' positions when nothing better exists rather than accept draw proposals.

Modelling that kind of human inability in terms of discounting the score seems to be somewhat of a quick and dirty solution, but finding a more appropriate model is difficult and probably of less interest to engine developers, as long as faster CPUs appear fairly regularly. If I remember, Will Hartston suggested -- possibly as a joke -- that play that requires activity on both flanks may be more difficult for some players to disentangle, than play 'in the same place'. Depth of action, together with 'bushiness' of movetree has also been suggested -- on rec.games.chess.computer, if I recall -- as a possible way to identify play that humans may not find easy to deal with, and so possibly 'contempt play': I'm going for complications in the firm belief your little mind can't cope with them. I don't expect to see you on the other side! Or 'thoughts' to that effect.

Skilled players can probably use configurable contempt factors to make their play more interesting, though.