Well, the Google AI Challenge is over, and I ended up as the second best Clojure bot. My skill rating was 53.23, which put me in 1,335^th place. I never expected to be so high, but it turns out there weren’t a lot of Clojure entries. I was beaten, by a large margin, by thobel. He did a great job. With the contest over, I thought I’d post a little postmortem of how my bot worked.

Global State & Exploration

My bot really doesn’t keep any global state, which is one of the reasons it didn’t do better. I planned to start adding some, but I never got around to it as I got distracted by work and other things. Here is the only global state that really matters:

(def ^{:dynamic true} *current-defense* (atom nil))
(def ^{:dynamic true} *positions-to-fill* (atom #{}))
(def ^{:dynamic true} *positions-unfilled* (atom {}))

(def ^{:dynamic true} *ant-last-moves* (atom {}))

The first three variables keep track of bot’s defenses, which is the only form of coordination between the ants. *ants-last-moves* keeps track of the last move each ant made so they can be prevented from backtracking.

In fact, the backtracking prevention is the only form of exploration in the code. Each ant sets out when spawned in a random direction, and ants have a 90% chance of continuing in the direction they were already traveling.

Core Logic

The core of the bot is a function called process-ants-for-moves. It is called once for each ant and is responsible for determining what that individual ant will do. In the last set of changes I uploaded, I added some code to keep track of how long the turns were taking and abort processing if it hit the 90% mark. It runs the process-ant function for each ant, building up a list of where the ants new positions are so they can be sent to the server at the end of processing.

The process-ant has a list of individual functions it can apply to the world state to find the best move for the ant. Cleaning things up with some pseudo code, it looks like this:

(defn find-move-through-functions
  "Run each function, return the first valid non-nil direction"
  [ant valid-directions valid-directions-with-back ants-last-move]
  (when (not-empty valid-directions))
    (loop [functions-to-run {move-to-defend-our-hills :defense
                            move-to-emergency-defense :e-defense
                            move-to-capture-hill :capture
                            move-away-from-enemy :run-away
                            move-towards-food-classic :food
                            move-in-random-direction :random}]
      (if (not-empty functions-to-run)   ; Still functions to check
        (let [the-function-stuff (first functions-to-run)
              the-function (key the-function-stuff)
              the-function-purpose (val the-function-stuff)
              result (apply the-function ...)]
          (if (= :none result)
            nil     ; The answer is 'don't move'
            (if (empty? result)
              (recur (rest functions-to-run)) ; No answer, try next
              (random direction from result))))))))

This works quite well. It runs each function in turn for the ant. If the function returns nil, the next function is run. Otherwise a random direction for the list of possibilities the function returns is chosen as the ant’s move.

The Move Functions

There are quite a few of these, including some from experimentation that don’t actually get run. Here is the description of those that are in use, in priority order. They all tend to follow the same pattern. I never put it into a macro or function, but that would have been smart. Here is the basic template:

(defn some-move-function
  [ant _ _]
  (cond
    (check that that function applies, such as if we see enemies)
      nil  ; It doesn't

    :else
      (let [distances (calculate distances to things of interest)
            spots (sort distances and filter out things too far away)
            visible-spots (stuff in spots that's line-of-site of ant)
            closest-spot (first (first visible-spots))]
        (when closest-spot    ; If something matches the criteria
          ; Some additional processing may be done here
          (utilities/direction ant closest-spot)))))

The line-of-sight code was a pretty big improvement. It meant that ants no longer got stuck against water trying to get to a piece of food on the other side. The function is basically Bresenham’s line algorithm. It’s not perfect, but it worked well enough and was pretty fast. As it walks along the line, it checks for water. If it doesn’t find any, the line-of-sight is clear.

Each function returns a set of directions the ant should move in, and the code picks one randomly. The selection is weighted so the ants tend to move in straight lines.

move-to-defend-our-hills

This is the function that uses up most of the global state. The global variable *current-defense* is checked to see what kind of defense we should be running. This is set at the start of each turn based on the number of ants we have per hill.

This is the code that sets up the little formations around my hills. First the ant is checked against the list of defense positions. If the ant is already on one of the spots, the answer is the keyword :none, meaning “don’t move from that spot.”

If the ant isn’t in a good position, it checks to see if it’s within visual range of a spot that does need filling. If so, it goes straight to it. This has the effect that newly spawned ants first action is to move into any holes in the defensive positions. Because of the way the defenses are setup (and since ants can’t move diagonally), attackers have to come in from a corner or two-abreast if they don’t want to lose their ants. This was remarkably effective.

move-to-emergency-defense

Once the defensive positions are filled, the next thing to do is to run emergency defenses. The code finds the closest visible hill to the ant, and then checks if any enemy ants are too close.

If they are, any ants that can see that hill rush back to try to stand on top of it. This floods the hill with defenders, hopefully allowing it to survive the onslaught. Since there is no real state on the ants, as soon as the danger is passed (the enemies are gone), they go straight back to normal behavior.

move-to-capture-hill

Whenever my ants see an enemy hill, they go straight for it. There is no hesitation, no strategy, just suicide charges. As a result of the next function (move-away-from-enemy), my ants don’t tend to get near enemy hills unless the area is not well defended. I’ll often lose ants this way, but it’s amazingly effective at grabbing hills people leave undefended. If an enemy moves so much as moves a square or two off a hill, my ants have a good chance at taking it.

move-away-from-enemy

This function is my bot’s entire survival instinct. My ants run in the other direction from the closest enemy that’s within 8 squares. This does hamper food gathering, but it prevents my ants from getting roundly slaughtered during movement. Note that this function doesn’t apply if the ant is in one of the defense modes (since this function wouldn’t get run), or when the ant is within view distance of one of my hills (an old condition from before defense mode so my ants wouldn’t give up their home bases).

move-towards-food-classic

It’s called “classic” because I came up with a better function for finding food, but it was never turned on due to performance problems. This function simply moves the ant towards the nearest piece of food they can see (with the line-of-sight check). In practice, having multiple ants go after the same piece of food wasn’t a problem.

moves-in-random-direction

If nothing else came up with a move, the ant would go in a random direction. This caused (limited) exploration.

Things I Meant To Do

If you look through the code, you can find commented out sections from some of my experiments. One of the early ones was move-towards-food-res. It would find the closest ant to each piece of food and record a reservation. This prevented multiple ants from going towards one piece of food. When I turned this on, my ants picked up much less food. The simpler method seemed to work better.

The better method of finding food was diffusion. There are globals and functions to take the food that the bot knows about and diffuse their influence, producing a gradient that the ants could follow to the areas with the most food (see influence mapping).

It turned out that this was much too slow. Others got it working (quite well), but I didn’t. Data structures were (I think) a big part of this. I was using a pseudo queue made out of a vector, and it wasn’t that fast. As it turns out Clojure has a queue that simply has no syntax yet which was very fast, but I didn’t get around to converting my code. The other thing that would have helped was not recalculating everything each turn. I started planning this but never got around to it either.

Many people used some form of this to explore the maps, which would have also been a smart thing to do.

Debugging

One thing I learned a fair bit about was debugging. It’s something difficult to do in Clojure (since there aren’t any native debuggers that I know about), so I had to invent some. I started debugging by sprinkling my code with statements like this:

(utilities/debug-log "Ant at " ant " doing " the-function-purpose ", going " dir)

The actual function looked like this:

(defn debug-log
  "Log something to the console for us to go through later"
  [& message]
  (when defines/logging-enabled
    (binding [*out* (if (nil? defines/*log-file*)
                        *err*
                        defines/*log-file*)]
      (apply println message)
      (flush))))

First the function checked to see if logging was enabled. If it was the *log-file* global variable was checked. When defined, the code would write the log message out to a file (which I would use to follow what my bot was doing). If the file wasn’t defined, it would just spit the message to standard error (which would cause the ants server to mark my bot as crashed for producing unexpected output).

Later I started using the visualizer that was available in the forums, but I had to be careful. I found the python server wouldn’t terminate if I produced too much visualization output.

Summary

That’s basically it. I’ve put my bot’s source up on my GitHub account if you want to take a look at it. The Clojure code is in the src/ directory, and it’s pretty clear that it was under active development when I stopped. Code is still commented out in places from debugging.

My ants had already learned courage, and now they know the importance of cowardice. It’s been about two weeks since I uploaded a new bot, and this new version is much improved. The last real version of my bot was able to play 63 games, and ended with a skill rating of about 58.

Today’s intelligence level: real swarm.

It took a couple of deploys to get the newest version of my bot up, but after only a few games it’s now rated at almost 69. At the moment I’m typing this, my bot is once again the best Clojure bot in the content, and is on the edge of the top 500 (although I expect to fall out shortly).

After doing some simple work on my bot (such as using the random seed the content provides to let you run game replays correctly), I found and fixed the bugs I had mentioned last time. The biggest problem the last version of my bot was having (besides stupid behavior due to bugs) was defense. Since the ants didn’t care too much about enemy ants (and sometimes actively avoid them) under the right circumstances it was possible to walk in and take one of my hills without too much effort. The bug fixes made this worse as the ants spread out better so there were fewer to act as defense.

This was fixed by implementing a strategy that I thought of a while ago, but was remedied of when I saw a competitor using it. By strategically placing ants around my hills I can raise the bar for anyone trying to attack. In the picture at the top of this post you can see my 2^nd level arrangement (out of three). This simple defense works great against single file streams of ants. Combined with new code that causes all nearby ants to run home if the hill gets attacked my hills have been able to survive much longer than they ever would have before. In some games, an enemy loses all their extra resources unsuccessfully trying to take my hill.

Things aren’t perfect. A side effect of the way things work is getting an ant near one of my hills can, under circumstances that aren’t nearly rare enough, prevent my bot from doing any further work. My colony may survive, but it’s not out gathering food or razing other hills. This in combination with a large number of opponent ants means that my bot can be shut down and eliminated by a correct show of force.

The last thing I had to do was fix a problem I didn’t expect to have: my bot times out. In easy games (especially the initial “does the bot crash” game bots run through when first uploaded) my bot can generate so many ants that it actually goes over it’s turn time. I added a little code so skip processing ants when my bot is about to run out of time, and I have seen it in action in production. Due to some quirk of the Python server implementation the order of ants to the bots being tested in is a discernible pattern. So when ants stop moving (because I ran out of time and didn’t get to issue them any orders) it makes a very obvious pattern on the field.

While the code seems to work very well, it does still fail on the ultra-small test map for some reason. I’m going to have to keep an eye on my bot to see if it happens in any other games. I may have to make it more aggressive.

My bot is getting better. After taking a little time off, I spent a few hours last night making updates. While I haven’t deployed them, my current bot has continued to test it’s potential. It’s been playing a lot of random_walk maps (which it handles better than the maze maps), and my rank is now in the top 1200.

Current intelligence level: stupid swarm.

While spending a bunch of time yesterday, I uncovered a couple of bugs in my code that explained some of the odd behavior I’ve seen (such as ants getting stuck together in small features). It turns out that I was using contains on a list, which does nothing. This is actually rather odd. contains checks to see if a key is in a collection, but doesn’t raise an exception in this circumstance. keys shows the behavior that I was expecting:

user=> (contains? (list 1 2 3 4) 3)
false
user=> (keys (list 1 2 3 4))
ClassCastException java.lang.Long cannot be cast to [...]

So because of this, some of the tests I had in my code were useless, essentially always returning false. On top of that, I found another bug. I was checking to see if collections were empty like this:

(if (empty my-collection) [...]

That’s a mistake too. empty is a new function in Clojure 1.3 which returns an empty version of the collection passed to it (so it will give you an empty set if you pass it a set). This also caused subtle bugs. The correct function was empty?, but I wasn’t looking that closely at autocomplete.

With that fixed, my bot looks quite a bit better. The ants don’t get stuck and they spread out better on all maps, even though there isn’t any code specifically instructing them to do that. When testing, I noticed that this means it’s much easier to take my hills because there are fewer ants milling about now. I added some code to have everyone rush home when an enemy approaches, but I haven’t given it any real test yet. It looked like it might have kicked in during  test game, but I was tired of working by then and decided to do the real testing later.

So as it is, I don’t want to post my updated version until I can test better. If I post what I have at this moment I’m pretty sure my rating will drop more than a few points. But I now know that my bot can handle over 600 ants without timing out. Before it almost never got above ~250 because the ants ended up blocking the hills.

Over the weekend I updated my bot to use a line of sight to decide on what to do for each ant. This fixes some of the worst behaviors my ants had, but it’s not perfect. The ants no longer crowd up trying to get to food or a hill they can’t reach, but they’re still not terribly bright. I had to turn off the food reservation system because it wasn’t intelligent enough and the side effects were worse than the benefits.

So how smart is my bot now? I’m not sure. It should be better, but the AI Challenge keeps having server problems. My bot had to wait most of the weekend for games to start being played again, and they seemed to have another problem today. Because of this games aren’t happening very fast at all so my bot hasn’t had the chance to move up to it’s true rank. The AI Challenge forms have posts about alternate servers you can test you bot on, I may have to go to that.

I’m a little tired of working on the bot at this point, even though I can think of quite a few things I should do. After all the time I’ve put in lately, it would be nice to see my bot reach it’s potential (or at least where it was before the last change). The slow game rate was very demotivating.

My bot topped out around 1000^th place. After a while it started to lose, which wasn’t that surprising. For various reasons, my bot is very bad on the maze maps and those seem to come up the majority of the time.

I spent the week trying to make my bot more intelligent. I’ve wanted to use gradients to give my bots marching orders, so I worked on code to do it. I wrote it over two days or so and rewrote it at least twice before I tested it as I kept thinking through the algorithm. I was really pleased when, except for one a mistyped variable name, it worked the first time.

I started testing how fast it was (since bots in the AI Challenge are limited to 3 seconds per turn) and it was great on tiny test cases. As I went to test it on random data that represented something the size of the large maps, it started crashing. I spent a bunch of time chasing down a StackOverflowError bug. I was pretty sure it was a bug in my code for quite a while, but as it turned out it’s just a known issue in Clojure, my code was technically correct. If you embed LazySeq in LazySeq in LazySeq over and over when Clojure goes to get a value it can go so deep that it runs out of stack space. The solution is to force evaluation to remove the “Lazy-ness”.

Unfortunately, things are still way too slow. Doing a full re-build every turn is never going to work, so I need to make it so I do partial updates every turn. While doing a little research I found that Clojure has a hidden queue type, which is drastically faster at removal (about 60x in a simple test case) than my List based solution.

In the mean time I have made my bot smarter. I’m hoping my bot is now up to tired toddler, but we’ll see as it starts to run games. I added a food reservation system to prevent tons of ants from getting piled up trying to get to a single piece. I also added some code to help with exploration, but I had to disable it as it caused a weird side effect I didn’t feel like debugging tonight. It actually caused more jams.

Speaking of jams, while watching my last replay I noticed an odd case. The ants are programed not to go backwards if possible. When a group of ants gets boxed in to a concave area in a map, if the outside ants push “in”, they won’t want to go backward and end up locking all the ants in. I did shuffle some code around, so this may no longer be an issue. I’ll just have to keep an eye on it.

I’ve uploaded a new version of my bot to the AI Challenge site. My bot was, I believe, the 10th best Clojure bot before I updated my code and wiped out my rating. The new code probably won’t play it’s first game until early tomorrow morning. I put in a fix for the possible issue of my ants failing to defend their hills, giving my ants courage when they could see one of my hills.

The problem with ants getting trapped that I mentioned before actually bit me. In my bot’s 5^th game, it’s first move was into such a corner. As a result, it did precisely nothing the entire match, until an enemy ant kindly came along and wiped me out.

At least that’s fixed now. New intelligence rating: very tired toddler.

My entry for Google’s Fall 2011 AI Challenge has been posted. In it’s first game it pulled out a decisive victory, in that it was gathering food but none of the other bots were. Still, it’s up there.

Current intelligence rating: badly sleep deprived toddler.

Right now, the logic is pretty simple. For each ant it works through a list of objectives, making the first move that an objective returns. It tries to raze opponent’s hills and gather food, all while running away from enemies. If it can’t do either of those, it will basically make a random move. There is code to prevent the ants from cycling between two squares, but it has an unintended consequence. The ants aren’t allowed to go back to the last square they were on. The leads to ants trapped on little islands surrounded by water (there are five such trapped ants in the shot above).

On the whole it works well. Since my colony spends so much time gathering food, it grows fast. The cowering I built into them helps keep them alive, and the random walking combined with the food seeking provides some exploration. In the end, the colony can grow pretty large and they tend to stay clumped up. I’d imagine I’m no match for the good bots, but I should be able to survive OK.

As I post this, there are only 25 Clojure based bots. I’m hoping to make a good showing on the list. That’s my main goal.