search

kkuette / TradzQAI

Trading environnement for RL agents, backtesting and training.
Apache License 2.0
165 stars 47 forks source link

Choosing different agent than PPO produces faulty behavior. #10

Closed oscarberonius closed 5 years ago

oscarberonius commented 5 years ago

I can run the default agent without issues, but had some trouble when running TRPO and DQN.

I tried removing both the save and config directories and rebuilding with trpo using 'python run.py -b TRPO'. The config directory seemed to be correctly created. The containing agent.json file had the correct property "type": "trpo_agent".

When I tried running the agent (python run.py) a save directory containing a PPO_0 directory was created. I tried changing the default agent in run.py to be TRPO as well but had the same result. The model could not evaluate and produced the following error.

Traceback (most recent call last):
  File "run.py", line 110, in <module>
    session.loadSession()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/TradzQAI/core/session/local.py", line 46, in loadSession
    self.initAgent()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/TradzQAI/core/session/local.py", line 57, in initAgent
    self.agent = self.agent(env=self.env, device=self.device)._get()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/TradzQAI/agents/PPO.py", line 6, in __init__
    Agent.__init__(self, env=env, device=device)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/TradzQAI/agents/agent.py", line 27, in __init__
    device=device
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/agents/agent.py", line 283, in from_spec
    kwargs=kwargs
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/util.py", line 192, in get_object
    return obj(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/agents/trpo_agent.py", line 161, in __init__
    entropy_regularization=entropy_regularization
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/agents/learning_agent.py", line 141, in __init__
    batching_capacity=batching_capacity
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/agents/agent.py", line 80, in __init__
    self.model = self.initialize_model()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/agents/trpo_agent.py", line 189, in initialize_model
    likelihood_ratio_clipping=self.likelihood_ratio_clipping
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/pg_prob_ratio_model.py", line 88, in __init__
    gae_lambda=gae_lambda
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/pg_model.py", line 98, in __init__
    requires_deterministic=False
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/distribution_model.py", line 90, in __init__
    discount=discount
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/memory_model.py", line 114, in __init__
    reward_preprocessing=reward_preprocessing
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/model.py", line 217, in __init__
    self.setup()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/model.py", line 290, in setup
    independent=independent
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/memory_model.py", line 605, in create_operations
    independent=independent
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/model.py", line 1195, in create_operations
    self.create_observe_operations(reward=reward, terminal=terminal)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/model.py", line 1129, in create_observe_operations
    reward=reward
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/memory_model.py", line 566, in tf_observe_timestep
    return tf.cond(pred=optimize, true_fn=true_fn, false_fn=tf.no_op)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py", line 488, in new_func
    return func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2076, in cond
    orig_res_t, res_t = context_t.BuildCondBranch(true_fn)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py", line 1920, in BuildCondBranch
    original_result = fn()
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/memory_model.py", line 562, in true_fn
    optimize = self.fn_optimization(**batch)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/pg_model.py", line 305, in tf_optimization
    next_internals=next_internals
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/models/memory_model.py", line 491, in tf_optimization
    return self.optimizer.minimize(**arguments)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/optimizer.py", line 139, in minimize
    deltas = self.step(time=time, variables=variables, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/optimized_step.py", line 102, in tf_step
    **kwargs
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/natural_gradient.py", line 123, in tf_step
    deltas = self.solver.solve(fn_x=fisher_matrix_product, x_init=None, b=[-grad for grad in loss_gradients])
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/solvers/conjugate_gradient.py", line 80, in tf_solve
    return super(ConjugateGradient, self).tf_solve(fn_x, x_init, b)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/solvers/iterative.py", line 78, in tf_solve
    args = tf.while_loop(cond=self.next_step, body=self.step, loop_vars=args)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py", line 3274, in while_loop
    return_same_structure)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2994, in BuildLoop
    pred, body, original_loop_vars, loop_vars, shape_invariants)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2929, in _BuildLoop
    body_result = body(*packed_vars_for_body)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 368, in __call__
    return self._call_func(args, kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/template.py", line 311, in _call_func
    result = self._func(*args, **kwargs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/solvers/conjugate_gradient.py", line 128, in tf_step
    A_conjugate = self.fn_x(conjugate)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorforce/core/optimizers/natural_gradient.py", line 112, in fisher_matrix_product
    return tf.gradients(ys=delta_kldiv_gradients, xs=variables)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py", line 596, in gradients
    gate_gradients, aggregation_method, stop_gradients)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py", line 749, in _GradientsHelper
    _RaiseNoGradWrtInitialLoopValError(op, from_ops, xs)
  File "/Users/Oscar/Documents/Cryptotrading/Tradz/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py", line 441, in _RaiseNoGradWrtInitialLoopValError
    % target_op.name)
ValueError: Cannot compute gradient inside while loop with respect to op 'trpo/actions-and-internals/layered-network/apply/dense0/apply/linear/apply/W'. We do not support taking the gradient wrt or through the initial value of a loop variable. Gradients can be computed through loop invariants or wrt the input parameters to the loop body.

Switching to dqn produced the same problems with the file structure, although the model could evaluate.

kkuette commented 5 years ago

I'm reworking the core of TradzQAI at this moment, this bug will be fixed soon.

kkuette commented 5 years ago

I've just pushed the reworked core, it should work now.