def __init__(self, world, agents, test_each=0, max_batch=16):
import numpy as np
if not isinstance(agents, list):
agents = [agents]
self.get_observation_shape = lambda: world.obs_shape
self.get_action_shape = lambda: world.act_shape
self.get_reward_shape = lambda: world.rew_shape
cache = []
def trajectories(agent, n):
trajs = world.trajectories(agent, n)
for t in trajs:
cache.append(t)
return trajs
self.trajectories = trajectories
def test(agent):
mandalka.evaluate(agent)
todo = test_each
while todo >= 1:
batch = min(max_batch, todo)
trajectories(agent, n=batch)
todo -= batch
for i, a in enumerate(agents):
test(Agent.build(a, self, i))
rng = np.random.RandomState()
def trajectories(_, n):
n = int(n)
assert n >= 1
idx = rng.choice(len(cache), size=n)
# TODO: this is unsafe, values could be modified outside
return [cache[i] for i in idx]
self.trajectories = trajectories
self.num_trajectories = lambda: len(cache)
