Description of the UCI protocol: https://ucichessengine.wordpress.com/2011/03/16/description-of-uci-protocol/
Let us parse the logs first:
In [1]:
%pylab inline
In [2]:
! grep "multipv 1" log3.txt | grep -v lowerbound | grep -v upperbound > log3_g.txt
In [3]:
def parse_info(l):
D = {}
k = l.split()
i = 0
assert k[i] == "info"
i += 1
while i < len(k):
if k[i] == "depth":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "seldepth":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "multipv":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "score":
if k[i+1] == "cp":
D["score_p"] = int(k[i+2]) / 100. # score in pawns
i += 3
elif k[i] == "nodes":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "nps":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "hashfull":
D[k[i]] = int(k[i+1]) / 1000. # between 0 and 1
i += 2
elif k[i] == "tbhits":
D[k[i]] = int(k[i+1])
i += 2
elif k[i] == "time":
D[k[i]] = int(k[i+1]) / 1000. # elapsed time in [s]
i += 2
elif k[i] == "pv":
D[k[i]] = k[i+1:]
return D
else:
raise Exception("Unknown kw")
In [4]:
# Convert to an array of lists
D = []
for l in open("log3_g.txt").readlines():
D.append(parse_info(l))
# Convert to a list of arrays
data = {}
for key in D[-1].keys():
d = []
for x in D:
if key in x:
d.append(x[key])
else:
d.append(-1)
if key != "pv":
d = array(d)
data[key] = d
The number of nodes searched depend linearly on time:
In [5]:
title("Number of nodes searched in time")
plot(data["time"] / 60., data["nodes"], "o")
xlabel("Time [min]")
ylabel("Nodes")
grid()
show()
So nodes per second is roughly constant:
In [6]:
title("Positions per second in time")
plot(data["time"] / 60., data["nps"], "o")
xlabel("Time [min]")
ylabel("Positions / s")
grid()
show()
The hashtable usage is at full capacity:
In [7]:
title("Hashtable usage")
hashfull = data["hashfull"]
hashfull[hashfull == -1] = 0
plot(data["time"] / 60., hashfull * 100, "o")
xlabel("Time [min]")
ylabel("Hashtable filled [%]")
grid()
show()
Number of nodes needed for the given depth grows exponentially, except for moves that are forced, which require very little nodes to search (those show as a horizontal plateau):
In [8]:
title("Number of nodes vs. depth")
semilogy(data["depth"], data["nodes"], "o")
x = data["depth"]
y = exp(x/2.2)
y = y / y[-1] * data["nodes"][-1]
semilogy(x, y, "-")
xlabel("Depth [half moves]")
ylabel("Nodes")
grid()
show()
In [9]:
title("Number of time vs. depth")
semilogy(data["depth"], data["time"]/60., "o")
xlabel("Depth [half moves]")
ylabel("Time [min]")
grid()
show()
In [10]:
title("Score")
plot(data["depth"], data["score_p"], "o")
xlabel("Depth [half moves]")
ylabel("Score [pawns]")
grid()
show()
Convergence of the variations:
In [11]:
for i in range(len(data["depth"])):
print "%2i %s" % (data["depth"][i], " ".join(data["pv"][i])[:100])