Investigate low-value-lambda behavior

[dimitry12]

[dimitry12]

Here score is average reward per episode over whole training period (100K environment steps), which is a good proxy:

• higher if trains faster
• higher if policy is stable