def test_stateful_agents():
"""
In the previous test all of the agents were static - e.g. calling them twice should give the same results
Here, we replace them with random agents with a specified seed so that we can identify if they are giving the
correct results.
"""
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 2.)
spec1 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2)
spec2 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=2, N=2)
agent1 = AgentFactory.create(spec1)
agent2 = AgentFactory.create(spec2)
agent1_duplicate = AgentFactory.create(spec1)
agent2_duplicate = AgentFactory.create(spec2)
comp_agent = AgentFactory.sequentialize([agent1_duplicate, agent2_duplicate], [3, 3])
delegates = [agent1, agent1, agent1, agent2, agent2]
for i in range(5):
check_same_behaviour(comp_agent, delegates[i], state, action)
comp_agent.view(AnonymizedHistoryItem())
delegates[i].view(AnonymizedHistoryItem())
def test_unequal_lengths():
"""
This is mostly just input checks.
The number of agents should be 1 more than the number of transitions
We make a number of calls where this relationship is violated and expect them to error.
"""
vote_bounds = (lambda x: 0., lambda x: 3.)
spec1 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2)
spec2 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=2, N=2)
spec3 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=3, N=2)
spec4 = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=4, N=2)
agent1 = AgentFactory.create(spec1)
agent2 = AgentFactory.create(spec2)
agent3 = AgentFactory.create(spec3)
agent4 = AgentFactory.create(spec4)
with pytest.raises(Exception):
AgentFactory.sequentialize([agent1, agent2], [1,2,3,4,5])
# too many transitions
with pytest.raises(Exception):
AgentFactory.sequentialize([agent1, agent2, agent3], [1, 2])
# too many agents
with pytest.raises(Exception):
AgentFactory.sequentialize([agent1, agent2, agent4], [1, 2, 3, 2])
def test_many_timesteps():
"""
Similar to above tests, except that we just run for an extra 1000 timesteps
after the last transition to make sure that it keeps using the last agent.
Returns
-------
"""
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 3.)
specs = [
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=2, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=3, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=4, N=2)]
agents = [AgentFactory.create(spec) for spec in specs]
duplicate_agents = [AgentFactory.create(spec) for spec in specs]
big_comp = AgentFactory.sequentialize(duplicate_agents, [2, 3, 4, 1])
delegates = ([agents[0]] * 2) + ([agents[1]] * 3) + ([agents[2]] * 4) + \
([agents[3]] * 1)
for i in range(10):
check_same_behaviour(big_comp, delegates[i % len(delegates)], state, action)
big_comp.view(AnonymizedHistoryItem())
delegates[i].view(AnonymizedHistoryItem())
def test_base_case():
"""
Simple base case test.
Creates static agents, put s them in a sequence, and checks that it gives the
correct output at each timestep.
Note that vote() and bet() are both called multiple times per timestep
"""
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
spec1 = AgentFactorySpec(AgentsEnum.constant, vote=3., prediction=[1, 2], bet=[.1, .2])
spec2 = AgentFactorySpec(AgentsEnum.constant, vote=4., prediction=[2, 2], bet=[.1, .0])
spec3 = AgentFactorySpec(AgentsEnum.constant, vote=2., prediction=[3, 1], bet=[.3, .1])
agent1 = AgentFactory.create(spec1)
agent2 = AgentFactory.create(spec2)
agent3 = AgentFactory.create(spec3)
comp_agent = AgentFactory.sequentialize([agent1, agent2, agent3], [2, 3, 4])
delegates = [agent1, agent1, agent2, agent2, agent2, agent3, agent3]
for i in range(7):
check_same_behaviour(comp_agent, delegates[i], state, action)
comp_agent.view(AnonymizedHistoryItem())
def test_recursive_case():
"""
Tests the case where some of the agents provided to sequentialize are themselves the result of sequentialize.
It's pretty likely that View() isn't being forwarded properly if this one doesn't work.
"""
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 3.)
specs = [
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=2, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=3, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=4, N=2)]
agents = [AgentFactory.create(spec) for spec in specs]
duplicate_agents = [AgentFactory.create(spec) for spec in specs]
comp1 = AgentFactory.sequentialize([duplicate_agents[0], duplicate_agents[1]], [3, 3])
comp2 = AgentFactory.sequentialize([duplicate_agents[2], duplicate_agents[3]], [3, 3])
big_comp = AgentFactory.sequentialize([comp1, comp2, duplicate_agents[0]], [5, 4, 2])
delegates = ([agents[0]] * 3) + ([agents[1]] * 2) + ([agents[2]] * 3) + \
([agents[3]] * 1) + ([agents[0]] * 2)
for i in range(10):
check_same_behaviour(big_comp, delegates[i], state, action)
big_comp.view(AnonymizedHistoryItem())
delegates[i].view(AnonymizedHistoryItem())
def test_negative_duration():
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 3.)
specs = [
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1,N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=2, N=2),
AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=0., seed=3, N=2)]
agents = [AgentFactory.create(spec) for spec in specs]
with pytest.raises(Exception):
AgentFactory.sequentialize(agents, [10, -1])
def test_super_long():
"""
Literally just a test with a bajillion agents.
Each agent lasts for 3 timesteps.
"""
N_AGENTS: int = 1000
FREQ: int = 3
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 3.)
spec = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2)
agents = [AgentFactory.create(spec) for _ in range(N_AGENTS)]
duplicate_agents = [AgentFactory.create(spec) for _ in range(N_AGENTS)]
transitions = [FREQ] * N_AGENTS
composite = AgentFactory.sequentialize(duplicate_agents, switch_at=transitions)
for i in range(FREQ * N_AGENTS):
delegated = agents[i // FREQ]
check_same_behaviour(composite, delegated, state, action)
composite.view(AnonymizedHistoryItem())
delegated.view(AnonymizedHistoryItem())
def test_duplicate_agents():
"""
Similar to previous tests, except that the same agent comes up multiple times in the agents list
TODO: test the case where agent functionality of the inner agent changes whenever View() is called
"""
env = EnvFactory.create(EnvsFactorySpec(EnvsEnum.default, n_actions=3))
state = env.state()
action = env.last_action
vote_bounds = (lambda x: 0., lambda x: 3.)
spec = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2)
agent = AgentFactory.create(spec)
duplicate_agent = AgentFactory.create(spec)
composite_agent = AgentFactory.sequentialize([agent] * 10, [2] * 10)
for _ in range(20):
check_same_behaviour(composite_agent, duplicate_agent, state, action)
composite_agent.view(AnonymizedHistoryItem())
duplicate_agent.view(AnonymizedHistoryItem())
def test_empty_arguments():
"""
Tests the case where at least one of agents or transitions has length 0 when being passed to
sequentialize()
"""
vote_bounds = (lambda x: 0., lambda x: 3.)
spec = AgentFactorySpec(AgentsEnum.random, bet=0.5, totalVotesBound=vote_bounds, vote=1., seed=1, N=2)
agent = AgentFactory.create(spec)
with pytest.raises(Exception):
AgentFactory.sequentialize([agent], [])
# it's possible this shouldn't raise an excpetion
with pytest.raises(Exception):
AgentFactory.sequentialize([], [1,])
# redundant with checking that the list is too long
with pytest.raises(Exception):
AgentFactory.sequentialize([], [])