Hi, when trying to run any of the provided examples, I get error messages regarding the "pickling" of an EvaluationTrue object such as :
Exception has occurred: TypeError
cannot pickle 'PyCapsule' object
File ".\msppy\quick_start\a_simple_MIP.py", line 47, in
resultTrue.run(n_simulations=100)

Is there something to resolve this?

Thanks and best regards
Joern

System: Windows 10, Python 3.10.5

As shown here, https://doi.org/10.1002/net.21932, a better way to do Periodical SDDP is to terminate the forward trial passes with the probability of 1 - the discount factor.

If you use a fixed number of stages for the forward pass, you will converge to a sub-optimal solution. This is particularly true if your initial x is far from the periodical optimum.

You can also use the probabilistically terminated forward passes as an upper bound.

I have a few queries regarding your implementation 😄

• Did you ever try any other methods apart from the level set for solving the SDDiP Lagrangian?
• Follow up: Do you ever notice bad numerics with SDDiP? i.e., cuts with very steep coefficients
• I see you implemented:

  msppy/msppy/msp.py

  Lines 782 to 787 in c708569

  	if var.X < var.lb:
  	solution[idx] = var.lb
  	elif var.X > var.ub:
  	solution[idx] = var.ub
  	else:
  	solution[idx] = var.X

  Did you implement any other tricks to improve numerics (SDDiP or otherwise)?
• I hadn't thought of adding a regularizer in the constraint:
  

  msppy/msppy/sp.py

  Lines 1386 to 1398 in 6e1d1de

  	self.rgl_constr = [self._model.addQConstr(
  	self.rgl -
  	a*gurobipy.QuadExpr(
  	gurobipy.quicksum([
  	self.states[i] * self.states[i]
  	- self.states[i] * 2 * center[i]
  	+ center[i] * center[i]
  	for i in range(self.n_states)
  	])
  	)
  	>=0,
  	name = 'rgl'
  	)]

  
  Did you do any benchmarking to see if it helps?