Hacker News Reader

Comment by danielatc | Hacker News Reader

The one thing that bugs me about the Google AI Challenges is that they do not really encourage using modern AI techniques. It's all about intelligent _developers_ rather than writing intelligent _software_. I really would like to see a challenge which is all about data: identifying patterns and learning to make predictions – rather than developing yet another heuristic for a minimax algorithm...

View on HN

I like that these AI Challenges appeal to everyone and rope in large numbers by the reward factor of watching your code fight someone else's to the death.

You're right, though, it's definitely not about writing intelligent software. That type of competition would have a much smaller but probably more intensely academic.

You are free to try to identify patterns in enemys ants behaviour, make predictions and win the game this way.

The challenge does not offer disk space for your bot to save data so it has to start from scratch every match (unless you supply it with a lot of offline training data).

There has been talk among the developers of the challenge of offering disk space in on of the future challenges.

An alternative would be to allow bots to connect to an external server maintained by the bot author to get the latest strategy, and dump match history for analysis.

Even just getting a second or two during startup to connect to external servers and again a second or two at the end would be enough.

This would be neat to do, but it introduces a number of additional implications. For instance, the scheduler must be considerably more fair in match allocation as this then has a huge impact on bot performance. It might also mean that newbie bots get trounced by more experienced bots so quickly that they don't learn much from the experience.

Not an easy problem, but perhaps a worthwhile one.

If I remember correctly, the last competition there was a team who wrote a bot using genetic programming techniques. So you are definitely not prevented from using those more advanced techniques if you can pull them off.

Food for thought, maybe you could develop an ML program to calculate the optimal heuristic for the basic bot program.

I think you would enjoy competing on a kaggle competition (google it).

http://www.kaggle.com/

Reply on news.ycombinator.com